DOMAIN GENERALIZABLE CONTINUAL LEARNING USING COVARIANCES

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 Amendment and Arguments Applicant’s amendment filed on January 2, 2026 has been entered and made of record. Claims 1-4, 7-13, and 16-20 are pending and are being examined in this application. In light of Applicant’s remarks and amendments to the claims, the 101 rejection is withdrawn. Applicant’s arguments with respect to 102 and 103 rejections have been considered, but are moot in view of the new ground(s) of rejection provided 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. Claims 1-4, 7, 9, 10-13, 16, 18, 19, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over LEE et al. (NPL “A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks”). Regarding claim 1, LEE teaches “A computer-implemented method for model training, comprising receiving, by a hardware processor, sets of images, each set corresponding to a respective task”: ([Page 2, Contribution, Paragraphs 1-2] “…Our high-level idea is to measure the probability density of test sample on feature spaces of DNNs (Deep Neural networks being the models to be trained) utilizing the concept of a “generative” (distance-based) classifier… We demonstrate the effectiveness of the proposed method using deep convolutional neural networks, such as DenseNet [14] and ResNet [12] trained for image classification tasks (a neural network classifier trained by receiving sets of images that are each based on various tasks such as CIFAR which contains object recognition tasks such as birds (https://www.cs.toronto.edu/~kriz/cifar.html) or SVHN, which contains number recognition tasks on street-view homes (https://www.kaggle.com/datasets/stanfordu/street-view-house-numbers)”) Additionally, neural networks are by definition, “computer program products” which by nature, computer-implemented. Further, LEE teaches “training, by the hardware processor, a task-based neural network classifier having a center and a covariance matrix for each of a plurality of classes in a last layer of the task-based neural network classifier and a plurality of convolutional layers preceding the last layer, by using a similarity between an image feature of a last convolutional layer from among the plurality of convolutional layers and the center and the covariance matrix for a given one of the plurality of classes, the similarity minimizing an impact of a data model forgetting problem”: ([Page 2, Contribution, Paragraphs 1-2] “…Our high-level idea is to measure the probability density of test sample on feature spaces of DNNs utilizing the concept of a “generative” (distance-based) classifier… Under this assumption, we define the confidence score using the Mahalanobis distance with respect to the closest class-conditional distribution (a well-known mathematical calculation that calculates similarity using the mean (center) and the covariance matrix), where its parameters are chosen as empirical class means and tied empirical covariance of training samples… We demonstrate the effectiveness of the proposed method using deep convolutional neural networks (neural networks with a plurality of convolutional layers), such as DenseNet [14] and ResNet [12] trained for image classification tasks (a neural network classifier trained using image features) on various datasets including CIFAR [15], SVHN [28], ImageNet [5] and LSUN [32]…”) And further: ([Pages 2-3, Why Mahalanobis distance-based score?] “Let PNG media_image1.png 15 48 media_image1.png Greyscale be an input and PNG media_image2.png 13 26 media_image2.png Greyscale PNG media_image3.png 18 102 media_image3.png Greyscale be its label. Suppose that a pre-trained softmax neural classifier is given (a pre-trained classifier based on Mahalanobis distance would then have a center and a covariance matrix for its known classes): PNG media_image4.png 30 205 media_image4.png Greyscale where wc and bc are the weight and the bias of the softmax classifier for class c, and PNG media_image5.png 16 26 media_image5.png Greyscale denotes the output of the penultimate layer of DNNs (the last layer preceded by a plurality of convolutional layers) … To estimate the parameters of the generative classifier from the pre-trained softmax neural classifier, we compute the empirical class mean and covariance of training samples PNG media_image6.png 17 141 media_image6.png Greyscale PNG media_image7.png 37 463 media_image7.png Greyscale where Nc is the number of training samples with label c. This is equivalent to fitting the class-conditional Gaussian distributions with a tied covariance to training samples under the maximum likelihood estimator. Mahalanobis distance-based confidence score. Using the above induced class-conditional Gaussian distributions, we define the confidence score M(x) using the Mahalanobis distance between test sample x and the closest class-conditional Gaussian distribution, i.e., PNG media_image8.png 22 278 media_image8.png Greyscale (Here, we see the Mahalanobis distance being used, which compares the similarities of a test sample (image feature) using the center and covariance to the known plurality of classes’ center and covariance matrices. Here, f(x) is a sample representation of the image feature. PNG media_image9.png 15 16 media_image9.png Greyscale is the class mean/center. Σ is the tied covariance estimated from the entire training set, which is explicitly multiplied with the squared difference vector via its inverse Σ-1 which is algebraically equivalent to computing a squared norm in a whitened space (i.e., multiplication by a decomposed covariance factor such as Σ-1/2) and thus the result, is an estimated curvature between the mean and the image feature.)”) And further: ([Page 4, Algorithm 1] “ PNG media_image10.png 199 531 media_image10.png Greyscale ”) This algorithm illustrates the process of computing the confidence score using Mahalanobis distance, covariance and mean (center) And further: ([Page 5, Algorithm 2] PNG media_image11.png 153 534 media_image11.png Greyscale ”) This algorithm illustrates the process when used with class-incremental learning, to minimize data forgetting, by comparing the similarity of each training sample of the preceding layer with the known classes. Further, LEE teaches “wherein training the neural network classifier minimizes a cross-entropy loss between a prediction and a class label”: ([Page 7, Comparison of robustness, sentence 8] “…we remark that our method using softmax neural classifier trained by standard cross entropy loss (cross-entropy is the method used, which is the standard method for neural networks, aiming to minimize the loss between predictions and class labels) typically outperforms the ODIN using softmax neural classifier trained by confidence loss…”) Further, LEE teaches “wherein training further comprises: receiving a new task to classify into at least one of a plurality of new classes; adding a … new center and covariance for the at least one of the plurality of new classes; and training the model using the new task to recognize the new task in the future”: ([Page 5, Algorithm 2] “ PNG media_image11.png 153 534 media_image11.png Greyscale ”) This shows the use of the methods within class-incremental learning where new tasks are classified into new classes. You can also see the new class mean (center) and the covariance for the new class explicitly being calculated here. Finally, updates the shared covariance and returns the mean and covariance of all classes, thereby training the model using the new task to recognize the new task in the future. LEE fails to explicitly teach that the new center is “randomly initialized.” However, YANG teaches that the new center is “randomly initialized”: ([Page 5, top half] “in step S2. The transformer state quantity setting clustering centre number, using non-supervised clustering method for clustering category classification of sample data after standardized processing specifically comprises: …the distance measurement function can also use the Mahalanobis distance … randomly selecting 6 sample data as the initial clustering centre vectors μ = (μ1, ..., μ6) by a distance function metric for each sample data with all the clustering centre vector μ i of the similarity according to the calculation result, the nearest center vector is determined as the type of the sample, until finishing all the sample classification; calculating the sample mean in each class, and as the new clustering centre vector μ * = (μ1 *, ..., μ6*), re-performing similarity calculation to all sample and category classification; repeating the calculating, when μ * = (μ1 *, ..., μ6 *) does not change, indicating that clustering is completed and stopping calculating, obtaining each sample category. It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to modify the classification taught by LEE so that updating the Mahalanobis distance-based classifier to include new classes includes randomly selecting initial clustering centre vectors when calculating the new class mean as taught by YANG, with a reasonable expectation of success. The motivation for doing so would have been to ensure the objectivity and rationality of the classification results (YANG [last paragraph of pg. 2]). Regarding claim 2, LEE teaches the limitations of claim 1. Further, LEE teaches “wherein the similarity is measured based on a distance calculation made using a Mahalanobis distance that induces a Riemmanian geometry”: ([Page 2, Contribution, Paragraph 1] “…Our high-level idea is to measure the probability density of test sample on feature spaces of DNNs utilizing the concept of a “generative” (distance-based) classifier… Under this assumption, we define the confidence score using the Mahalanobis distance with respect to the closest class-conditional distribution (a well-known mathematical calculation used in the last layer that calculates similarity using the mean (center) and the covariance matrix), where its parameters are chosen as empirical class means and tied empirical covariance of training samples…”) This shows that the similarity measure/confidence score is based on a distance calculation made using a Mahalanobis distance. And further: ([Page 5, Algorithm 2] PNG media_image11.png 153 534 media_image11.png Greyscale ”) The above equations show a covariance matrix Σ that varies smoothly, thus inducing a non-flat Riemannian metric / a Riemannian geometry. Regarding claim 3, LEE teaches the limitations of claim 2. Further, LEE teaches “wherein the distance calculation for a given one of the plurality of classes is a squared distance of a difference between a sample representation of the image feature and a class-center of the given one of the plurality of classes multiplied by a decomposed covariance”: ([Page 3, Equation 2] “ PNG media_image8.png 22 278 media_image8.png Greyscale ”) Here, f(x) is a sample representation of the image feature. PNG media_image9.png 15 16 media_image9.png Greyscale is the class center. The subtraction and squaring via quadratic form matches the claimed “squared distance of a difference between a sample representation of the image feature and a class-center of the given one of the plurality of classes”, where Σ is the tied covariance estimated from the entire training set, which is explicitly multiplied with the squared difference vector via its inverse Σ-1 which is algebraically equivalent to computing a squared norm in a whitened space (i.e., multiplication by a decomposed covariance factor such as Σ-1/2) Regarding claim 4, LEE teaches the limitations of claim 3. Further LEE teaches “wherein the distance calculation is a prediction”: ([Page 4, Paragraph 1] “We remark that this corresponds to predicting a class label using the posterior distribution from generative classifier with the uniform class prior. Interestingly, we found that the softmax accuracy (red bar) is also achieved by the Mahalanobis distance-based classifier (blue bar), while conventional knowledge is that a generative classifier trained from scratch typically performs much worse than a discriminative classifier such as softmax.”) This citation simply shows that a classifying data with a neural network classifier inherently corresponds with making a prediction, and thus, using a Mahalanobis distance to train the classifier is a prediction. Regarding claim 7, LEE teaches the limitations of claim 1. Further LEE teaches “wherein the neural network is trained using training data comprising respective pluralities of images pertaining to respective given tasks with distinct classes and domains”: ([Page 2, Contribution, paragraph 2] “We demonstrate the effectiveness of the proposed method using deep convolutional neural networks, such as DenseNet [14] and ResNet [12] trained for image classification tasks on various datasets including CIFAR [15], SVHN [28], ImageNet [5] and LSUN [32]…”) Here, we see the datasets used as training data, including CIFAR, which comprises 10 classes of a specific domain (natural images) (https://www.cs.toronto.edu/~kriz/cifar.html), the SVHN dataset, which comprises 10 classes of a different domain (street numbers of houses) (https://www.kaggle.com/datasets/stanfordu/street-view-house-numbers) in addition to ImageNet and LSUN, thus providing evidence of training data comprising pluralities of images pertaining to respective given tasks (number recognition, object recognition such as birds or boats) with distinct classes and domains. Regarding claim 9, LEE teaches the limitations of claim 1. Further LEE teaches “wherein the task-based neural network classifier uses covariance to estimate a curvature between a mean in the task-based neural network classifier and an image feature from a new task”: ([Page 2, Contribution, Sentence 3] “Given deep neural networks (DNNs) with the softmax classifier, we propose a simple yet effective method for detecting abnormal samples such as out of-distribution (OOD) and adversarial ones. We first present the proposed confidence score based on an induced generative classifier under Gaussian discriminant analysis (GDA) (it can be understood by one skilled in the art that Gaussian discriminant analysis is a known technique that uses a covariance matrix to explicitly model the curvature of the feature distribution around the mean (center)), and then introduce additional techniques to improve its performance. We also discuss how the confidence score is applicable to incremental learning.”) And further: ([Page 3, Equation 2] “ PNG media_image8.png 22 278 media_image8.png Greyscale ”) Here, f(x) is a sample representation of the image feature. PNG media_image9.png 15 16 media_image9.png Greyscale is the class mean/center. Σ is the tied covariance estimated from the entire training set, which is explicitly multiplied with the squared difference vector via its inverse Σ-1 which is algebraically equivalent to computing a squared norm in a whitened space (i.e., multiplication by a decomposed covariance factor such as Σ-1/2) and thus the result, is an estimated curvature between the mean and the image feature. Regarding claim 10, see at least the rejection for claim 1. LEE further teaches A computer program product for model training, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform the claimed method: ([Page 1, Abstract]) Neural networks are, by definition, “computer program products,” which are computer-executed. Regarding claim 11, see the rejection for claim 2. Regarding claim 12, see the rejection for claim 3. Regarding claim 13, see the rejection for claim 4. Regarding claim 16, see the rejection for claim 7. Regarding claim 17, see the rejection for claim 8. Regarding claim 18, see the rejection for claim 9. Regarding claim 19, see at least the rejection for claim 1. LEE further teaches A computer processing system for model training, comprising: a memory device for storing program code; and a hardware processor operatively coupled to the memory device for running the program code to perform the claimed steps: ([Page 1, Abstract]) Neural networks are, by definition, “computer program products,” which are executed by a computer comprising at least a hardware processor and a memory device. Regarding claim 20, see the rejection for claim 2. Claims 8 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over LEE and YANG, as applied to claims above, and further in view of TARVAINEN et al. (NPL “Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results”). Regarding claim 8, LEE and YANG teach the limitations of claim 1. LEE and YANG fail to explicitly teach “calculating a knowledge distillation loss by calculating a smooth transition coefficient between a current task-based neural network classifier and a prior task-based neural network classifier for a given task and further calculating an exponential moving average.” However, analogous art, TARVAINEN, does teach this: ([Pages 2-3, Section 2. Mean Teacher] “To overcome the limitations of Temporal Ensembling, we propose averaging model weights instead of predictions. Since the teacher model (a prior task-based neural network classifier) is an average of consecutive student models, we call this the Mean Teacher method (Figure 2). Averaging model weights over training steps tends to produce a more accurate model than using the final weights directly [19]. We can take advantage of this during training to construct better targets. Instead of sharing the weights with the student model (a current task-based neural network classifier), the teacher model uses the EMA weights of the student model (explicit calculation of an exponential moving average). Now it can aggregate information after every step instead of every epoch. In addition, since the weight averages improve all layer outputs, not just the top output, the target model has better intermediate representations. These aspects lead to two practical advantages over Temporal Ensembling: First, the more accurate target labels lead to a faster feedback loop between the student and the teacher models, resulting in better test accuracy. Second, the approach scales to large datasets and on-line learning. More formally, we define the consistency cost J (the knowledge distillation loss) as the expected distance between the prediction of the student model (with weights PNG media_image12.png 14 9 media_image12.png Greyscale and noise PNG media_image13.png 13 7 media_image13.png Greyscale ) and the prediction of the teacher model (with weights PNG media_image14.png 16 11 media_image14.png Greyscale and noise PNG media_image15.png 17 12 media_image15.png Greyscale ). PNG media_image16.png 29 260 media_image16.png Greyscale The difference between the II model, Temporal Ensembling, and Mean teacher is how the teacher predictions are generated. Whereas the II model uses PNG media_image14.png 16 11 media_image14.png Greyscale = PNG media_image12.png 14 9 media_image12.png Greyscale , and Temporal Ensembling approximates PNG media_image17.png 16 66 media_image17.png Greyscale with a weighted average of successive predictions, we define PNG media_image18.png 17 15 media_image18.png Greyscale at training step t as the EMA (explicit calculation of expected moving average) of successive PNG media_image12.png 14 9 media_image12.png Greyscale weights: PNG media_image19.png 19 143 media_image19.png Greyscale where PNG media_image20.png 12 11 media_image20.png Greyscale is a smoothing coefficient hyperparameter ( a smooth transition coefficient). An additional difference between the three algorithms is that the II model applies training to PNG media_image14.png 16 11 media_image14.png Greyscale whereas Temporal Ensembling and Mean Teacher treat it as a constant with regards to optimization.”) Here, the “consistency cost”, which is calculated as the mean squared error between teacher and student predictions, is equivalent to the claimed “knowledge distillation loss” calculated from “a smooth transition coefficient” ( PNG media_image20.png 12 11 media_image20.png Greyscale ) between a current task-based neural network classifier (student model PNG media_image21.png 17 11 media_image21.png Greyscale ) and a prior task-based neural network classifier (teacher model PNG media_image18.png 17 15 media_image18.png Greyscale ) It would have been obvious to one of ordinary skill in the art, prior to the effective filing date of the claimed invention, to modify the classification taught by the combination of LEE and YANG to include the calculation of a consistency cost taught by TARVAINEN, with a reasonable expectation of success. The motivation for doing so would have been because, as TARVAINEN points out, ([Section 2. Mean Teacher] “These aspects lead to two practical advantages over Temporal Ensembling: First, the more accurate target labels lead to a faster feedback loop between the student and the teacher models, resulting in better test accuracy. Second, the approach scales to large datasets and on-line learning”.) Regarding claim 17, see the rejection for claim 8. 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. Contact Information Any inquiry concerning this communication or earlier communications from the examiner should be directed to GRACE PARK whose telephone number is (571)270-7727. The examiner can normally be reached M-F 8AM-5PM. 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, TAMARA KYLE can be reached at (571)272-4241. 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. /Grace Park/Primary Examiner, Art Unit 2144