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.
Claim(s) 1-2 and 4-5 is/are rejected under 35 U.S.C. 103 as being unpatentable over Franceschi et al. (NPL: Unsupervised Scalable Representation Learning for Multivariate Time Series, published 2019, hereinafter “Franceschi”) in view of Ton et al. (NPL: Noise Contrastive Meta-Learning for Conditional Density Estimation using Kernel Mean Embeddings, published Apr. 2021, hereinafter “Ton”).
Regarding claim 1, Franceschi teaches a training method performed through batch learning by a computer, the training method comprising:
obtaining training data including first time-series data and second time-series data different from the first time-series data (Franceschi, Section 3 Paragraph 3 – “To adapt this principle to time series, we consider (see Figure 1 for an illustration) a random subseries2 xref of a given time series yi. Then, on one hand, the representation of xref should be close to the one of any of its subseries xpos (a positive example). On the other hand, if we consider another subseries xneg (a negative example) chosen at random (in a different random time series yj if several series are available, or in the same time series if it is long enough and not stationary), then its representation should be distant from the one of xref. Following the analogy with word2vec, xpos corresponds to a word, xref to its context, and xneg to a random word. To improve the stability and convergence of the training procedure as well as the experimental results of our learned representations, we introduce, as in word2vec, several negative samples (xneg k )k∈1,K , chosen independently at random.” and Fig. 1 – teaches obtaining training data including first time-series data and second time-series data different from the first time series data (Fig. 1 shows first and second time series data, time-series y(l, i, j), and sub-series x);
performing, using a contrastive learning algorithm, second training processing (Franceschi, Section 3 Paragraph 1 – “To this end, we introduce a novel triplet loss for time series, inspired by the successful and by now classic word representation learning method known as word2vec (Mikolov et al., 2013). The proposed triplet loss uses original time-based sampling strategies to overcome the challenge of learning on unlabeled data.”, Section 3 Paragraph 3 – “the representation of xref should be close to the one of any of its subseries xpos (a positive example). On the other hand, if we consider another subseries xneg (a negative example) chosen at random (in a different random time series yj if several series are available, or in the same time series if it is long enough and not stationary), then its representation should be distant from the one of xref.”, Eq. (1), and in Section 3 Paragraph 4 – “Overall, the training procedure consists in traveling through the training dataset for several epochs (possibly using mini-batches), picking tuples xref, xpos,(xneg k )k at random as detailed in Algorithm 1, and performing a minimization step on the corresponding loss for each pair, until training ends.” – teaches performing training using a contrastive learning algorithm (Eq. 1 describes a contrastive learning algorithm with triplet loss, wherein samples of the same time-series are brought closer to each other as positive samples and samples of different time series are pushed away from each other as negative samples)) of
(i) training the neural process model to bring first sampling data items close to each other as positive samples, the first sampling data items being generated by sampling from the first time-series data distribution, (ii) training the neural process model to bring second sampling data items close to each other as positive samples, the second sampling data items being generated by sampling from the second time-series data distribution (Franceschi, Section 3 Paragraph 3 – “To adapt this principle to time series, we consider (see Figure 1 for an illustration) a random subseries2 xref of a given time series yi. Then, on one hand, the representation of xref should be close to the one of any of its subseries xpos (a positive example). On the other hand, if we consider another subseries xneg (a negative example) chosen at random (in a different random time series yj if several series are available, or in the same time series if it is long enough and not stationary), then its representation should be distant from the one of xref. Following the analogy with word2vec, xpos corresponds to a word, xref to its context, and xneg to a random word. To improve the stability and convergence of the training procedure as well as the experimental results of our learned representations, we introduce, as in word2vec, several negative samples (xneg k )k∈1,K , chosen independently at random.” and Fig. 1 – teaches training the neural process model to bring first sampling data items close to each other as positive samples, the first sampling data items being generated by sampling from the first time-series data distribution (samples xref, independently and randomly, from series y where the representation of xref should be close to the one of any of its subseries xpos providing a positive example), and training the neural process model to bring second sampling items close to each other as positive samples, the second sampling items being generated by sampling from the second-time series distribution (samples selected from the same sub-series are pulled closer together, providing positive examples, thus second samples from a second series y are brought closer together)), and
(iii) training the neural process model to push away the first sampling data items and the second sampling data items far from each other as negative samples (Franceschi, Section 3 Paragraph 3 – “To adapt this principle to time series, we consider (see Figure 1 for an illustration) a random subseries2 xref of a given time series yi. Then, on one hand, the representation of xref should be close to the one of any of its subseries xpos (a positive example). On the other hand, if we consider another subseries xneg (a negative example) chosen at random (in a different random time series yj if several series are available, or in the same time series if it is long enough and not stationary), then its representation should be distant from the one of xref. Following the analogy with word2vec, xpos corresponds to a word, xref to its context, and xneg to a random word. To improve the stability and convergence of the training procedure as well as the experimental results of our learned representations, we introduce, as in word2vec, several negative samples (xneg k )k∈1,K , chosen independently at random.” and Fig. 1 – teaches training the neural process model to push away the first sampling data items and the second sampling data items far from each other as negative samples (when considering another subseries xneg from a different random time series y then its representation should be distant from xref, thus pushing away first sampling data items from second sampling data items as negative samples)),
wherein:
the neural process model includes an encoder that outputs feature representations in a latent space from time-series data (Franceschi, Section 3 Paragraph 3 – “To adapt this principle to time series, we consider (see Figure 1 for an illustration) a random subseries xref of a given time series yi.”, Eq. (1), and in Section 3 Paragraph 4 – “This loss pushes the computed representations to distinguish between xref and xneg, and to assimilate xref and xpos… The overall computational and memory cost is O(K ·c(f)), where c(f) is the cost of evaluating and backpropagating through f on a time series; thus this unsupervised training is scalable as long as the encoder architecture is scalable as well.” – teaches wherein the model includes an encoder that outputs feature representations in a latent space from time-series data (uses an encoder to compute representations from time-series data for the loss function that pushes computed representations to distinguish between xref and xneg, thus the model includes an encoder that outputs feature representations in a latent space from time-series data));
in the second training processing, a processing target of the training performed using the contrastive learning algorithm is the feature representations output from the encoder (Franceschi, Eq. (1) and in Section 3 Paragraph 4 – “This loss pushes the computed representations to distinguish between xref and xneg, and to assimilate xref and xpos… The overall computational and memory cost is O(K ·c(f)), where c(f) is the cost of evaluating and backpropagating through f on a time series; thus this unsupervised training is scalable as long as the encoder architecture is scalable as well.” – teaches wherein in the second training processing, a processing target of the training performed using the contrastive learning algorithm (algorithm of Eq. 1) is the feature representations output from the encoder (this loss pushes the computed representations of the encoder to distinguish between xref and xneg, thus a processing target of the training performed using the contrastive learning algorithm is the feature representations output from the encoder));
a parameter trained in the second training processing is a parameter of the encoder (Franceschi, Section 3 Paragraph 4 – “The objective corresponding to these choices to minimize during training can be thought of the one of word2vec with its shallow network replaced by a deep network f(·,θ) with parameters θ, or formally Eq. (1)” and in Section 3 Paragraph 5 – “We highlight that this time-based triplet loss leverages the ability of the chosen encoder to take as input time series of different lengths. By training the encoder on a range of input lengths going from one to the length of the longest time series in the train set, it becomes able to output meaningful and transferable representations” – teaches wherein a parameter (such as parameters θ in Eq. (1)) in the second training processing is a parameter of the encoder (uses triplet loss, or contrastive loss, to train encoder, thus a parameter trained in the second training processing is a parameter of the encoder)).
Franceschi fails to explicitly teach performing first training processing of training a neural process model to predict, based on the first time-series data and the second time-series data, a first time-series data distribution indicating a statistical characteristic of the first time-series data and a second time-series data distribution indicating a statistical characteristic of the second time-series data, the neural process model being a deep learning model that outputs, using a stochastic process, a prediction result that takes uncertainty into account.
However, analogous to the field of the claimed invention, Ton teaches:
performing first training processing of training a neural process model to predict, based on the first time-series data and the second time-series data, a first time-series data distribution indicating a statistical characteristic of the first time-series data and a second time-series data distribution indicating a statistical characteristic of the second time-series data, the neural process model being a deep learning model that outputs, using a stochastic process, a prediction result that takes uncertainty into account (Ton, Section 2.1 Paragraph 1 – “Formally, given sets X and Y, with a distribution P over the random variables (X,Y ) taking values in X × Y, the conditional mean embedding (CME) of the conditional distribution of Y |X = x, assumed to have density p(y|x), and is defined as: Eq. (1)” and in Section 2.1 Paragraph 2 – “Intuitively, Eq.1 allows us to represent a probability distribution p(y|x) in a function space (RKHS), by taking the expectation of features φy(y) ∈ HY under p(y|x).” and in Section 2.1 Last Paragraph – “Now that we can compute the CME for a given new x, i.e. embed the conditional distribution p(y|x) into a RKHS, the problem is how we are going to model the conditional density using the CME. In order to do that, we use ideas from noise contrastive estimation.” – teaches performing a first training processing of training a neural process model to predict, based on the first data and second data, a data distribution indicating a statistical characteristic of the first data and a second distribution indicating a statistical characteristic of the second data (finds a distribution indicating a statistical characteristics of any data set, given two sets X and Y), the neural process model being a deep learning model (embeds distribution into a reproducing kernel Hilbert space for modeling conditional density) that outputs, using a stochastic process (distribution over random variables, thus using a stochastic process), a prediction result that takes uncertainty into account (prediction result using stochastic process to produce a distribution of the given sets and model conditional density which takes uncertainty into account)); and
performing, using a contrastive learning algorithm, second training processing (Ton, Section 2.1 Last Paragraph – “Now that we can compute the CME for a given new x, i.e. embed the conditional distribution p(y|x) into a RKHS, the problem is how we are going to model the conditional density using the CME. In order to do that, we use ideas from noise contrastive estimation.” – teaches performing second training processing using a contrastive learning algorithm),
wherein:
the neural process model includes an encoder that outputs feature representations in a latent space (Ton, Fig. 1 and Section 3.2 Paragraph 6 – “We note that the concatenation of x and y encodes the joint distribution, rather than the conditional as in MetaCDE. While such task representation does preserve the relevant information, it is susceptible to changes in the marginal of x across tasks and conditional representations are intuitively better suited for the task of conditional density estimation.” – teaches wherein the neural process model includes an encoder that outputs feature representations (concatenation of x and y encodes joint distribution to output feature representations) in a latent space (as in Fig. 1, data is passed through feature maps to construct conditional mean embedding operator));
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the neural process models and first training processing to predict first and second distributions and output accounting for uncertainty of Ton to the second training processing, time-series data, encoders, and contrastive learning algorithm of Franceschi. Doing so would allow neural process models to be applicable to much broader sets of relationships between random objects, enable training systems on unseen and small datasets to estimate conditional densities, and utilize contrastive learning to model densities (Ton, Introduction).
Claim 4 and 5 incorporate substantively all the limitations of claim 1 in a device and a non-transitory computer-readable medium, and are rejected on similar grounds as above. Franceschi teaches the processors, memory, and medium of these claims in Section 5 Paragraph 1 – “Each encoder was trained using the Adam optimizer (Kingma & Ba, 2015) on a single Nvidia Titan Xp GPU with CUDA9.0, unless stated otherwise.”
Regarding claim 2, the combination of Franceschi and Ton teaches the training method according to claim 1, wherein:
the first time-series data is time-series sampling data obtained by sampling temporally-continuous first data[[,]]; and the second time-series data is time-series sampling data obtained by sampling temporally-continuous second data (Franceschi, Section 3 Paragraph 3 – “To adapt this principle to time series, we consider (see Figure 1 for an illustration) a random subseries xref of a given time series yi. Then, on one hand, the representation of xref should be close to the one of any of its subseries xpos (a positive example). On the other hand, if we consider another subseries xneg (a negative example) chosen at random (in a different random time series yj if several series are available, or in the same time series if it is long enough and not stationary)” and Fig. 1 – teaches wherein the first time-series data is time-series sampling obtained by sampling-temporally continuous first data, and the second time-series data is time-series sampling data obtained by sampling temporally-continuous second data (Fig. 1 shows samples x being sampled from temporally-continuous first and second data y(i, j))).
Claim(s) 3 is/are rejected under 35 U.S.C. 103 as being unpatentable over Ton and Franceschi as applied to claims 1 and 4-5 above, and further in view of Yang et al. (US Pub. No. 2021/0064931, filed Aug. 2020, hereinafter “Yang”).
Regarding claim 3, the combination of Franceschi and Ton teaches the training method according to claim 1, wherein:
the performing of the first training processing and the second training processing includes using an error function in which a second error function is changed by adding a term of a first error function used in the contrastive learning algorithm to a term of the second error function, the first error function reducing an error in a case of the positive samples and increasing the error in a case of the negative samples, the second error function pertaining to an error of a prediction result used by the neural process model (Ton, Section 3.3 Paragraph 1 – “We can now train the classifier using the model given in Eq.6 by maximizing conditional loglikelihood of the True/Fake labels, or equivalently Eq. (11). After parameters θ of the classifier have been learned, the conditional density estimates can be read off from Eq.4.” – teaches wherein performing of the first training processing and the second training processing includes using an error function in which a second error function is changed by adding a term of a first error function used in the contrastive learning algorithm to a term of the second error function (in Eq. (11), the contrastive loss term is added to an error of a prediction result, thus the second error function is changed by adding a term of the first error function), the first error function reducing an error in a case of the positive samples and increasing the error in a case of the negative samples (in Eq. (11), the second term corresponds to the contrastive loss term of Eq. (6)), the second error function pertaining to an error of a prediction result used by the neural process model (in Eq. (11), the first term corresponds to an error of a prediction result used by the neural process model, specifically corresponding to the conditional density estimates of Eq. (4))).
The combination of Ton and Franceschi fails to explicitly teach wherein the first training processing and the second training processing are performed concurrently.
However, analogous to the field of the claimed invention, Yang teaches:
wherein the first training processing and the second training processing are performed concurrently (Yang, Fig. 2A and [0057] – “One use of the learned hierarchical motion features is for improving video action recognition. To integrate the learned motion features into a backbone network, the prime motion block may be wrapped into a residual block: .sup.l=.sup.l+g.sup.l(.sup.l), where .sup.l is the convolutional features at level l, .sup.l is the corresponding motion features obtained as described above, and g.sup.t(.Math.) is a 1×1×1 convolution. This seamless integration enables end-to-end fusion of appearance and motion information over multiple levels throughout a single unified network, instead of learning them disjointly like two-stream networks. After the motion representations are self-supervised learned at all levels, the classification loss is added to jointly optimize the total objective, which is a weighted sum of the following losses shown in Equation 5.” – teaches wherein the first training processing and the second training processing are performed concurrently (jointly optimizes a first training process and second training process, where a prediction and contrastive learning process are performed concurrently))[[,]];
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the concurrent training processing of Yang to the first and second training processing of Ton and Franceschi. Doing so would enable the process model to jointly optimize the total objective through a unified network, instead of disjoint learning (Yang, [0057]).
Response to Arguments
Applicant’s arguments, see pp. 2 of Remarks, filed 24 July 2026, with respect to Claim Interpretation under 112(f) have been fully considered and are persuasive. The prior 112(f) interpretation of the limitation of claim 4 is withdrawn.
Applicant's arguments, see pp. 2-4 of Remarks, filed 24 July 2026, with respect to Claim Rejections under 35 U.S.C. 103 have been fully considered but they are not persuasive. Franceschi teaches the amended limitations of claim 1 regarding “the neural process model includes an encoder…”, “in the second training processing, a processing target of the training performed…”, and “a parameter trained in the second training processing is…”. Ton teaches the amended limitations of claim 1 regarding “the neural process model includes an encoder…”.
Applicant argues on pp. 3 that Franceschi fails to apply a contrastive learning algorithm to latent-space feature representations output from an encoder. Examiner respectfully disagrees. Franceschi at Eq. (1) and in Section 3 Paragraph 4 – teaches applying the contrastive learning algorithm (algorithm of Eq. 1) to latent-space feature representations output from the encoder (loss pushes the computed representations of the encoder to distinguish between xref and xneg, thus a processing target of the training performed using the contrastive learning algorithm is the feature representations output from the encoder).
Applicant further argues on pp. 3 of Remarks that Ton fails to employ a neural process model encoder. Examiner respectfully disagrees. Ton at Fig. 1 and Section 3.2 Paragraph 6 – teaches wherein the neural process model includes an encoder that outputs feature representations (concatenation of x and y encodes joint distribution to output feature representations) in a latent space (as in Fig. 1, data is passed through feature maps to construct conditional mean embedding operator). Ton further states at Section 1 Paragraph 3 that the methods and models contributed parallel that of Neural Processes and conditional Neural Processes.
In response to applicant's arguments against the references individually, one cannot show nonobviousness by attacking references individually where the rejections are based on combinations of references. See In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981); In re Merck & Co., 800 F.2d 1091, 231 USPQ 375 (Fed. Cir. 1986).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Xie et al. (US Pub. No. 2019/0095464, published Sept. 2018) teaches a dual deep learning architecture for machine learning systems with a relationship and embedding subnetwork. Teaches wherein the embedding subnetwork may be trained using a contrastive loss and wherein the relationship subnetwork may be trained using a predictive loss.
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 LOUIS C NYE whose telephone number is 571-272-0636. The examiner can normally be reached Monday - Friday 9:00AM - 5:00PM.
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, MATT ELL can be reached at 571-270-3264. 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.
/LOUIS CHRISTOPHER NYE/Examiner, Art Unit 2141
/MATTHEW ELL/Supervisory Patent Examiner, Art Unit 2141
Prosecution Insights