DETAILED ACTION
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA . This action is responsive to claims filed 5/11/2026. Claims 1-4, 7-14, 17-21 are pending.
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) 1-3, 9, 11-13, 19, 21 are rejected under 35 U.S.C. 103 as being unpatentable over Lewis (US 20220237465 A1) in view of Kingma ("Variational dropout and the local reparameterization trick", published 2015).
For claim 1, Lewis discloses: a computer-implemented (fig.8 shows computing device overview) method of generating a machine learning model for classifying sequential data (fig.1, 0027-29 contemplates various data including sequential data such as text, video, audio, sensor, etc.), the computer-implemented method comprising:
receiving the sequential data, wherein the sequential data comprises a plurality of records having a plurality of features in a plurality of environments (ibid: sequential data encoded into vector form, hence, plurality of features; the cited passages also disclose a plurality of environments for each mode (e.g., different vehicle environments for automatic vehicle navigation, field data for forms, topics, etc.) as well as the modes themselves being different environments; see also 0061: training data from different scene environments);
initializing a set of mask weights on the plurality of features (0044-46: parameter tensors corresponding to divided volumes or sparsity constraints constitute pruning masks on the features / intermediate layer weights; see also 0063-66: masking weights by a product between a scalar tensor and parameter tensors, hence, stochastic parameter tensors being a mask on the weights);
initializing a set of classifier weights on the plurality of features, wherein each mask weight of the set of mask weights corresponds to one classifier weight of the set of classifier weights, and each mask weight influences a corresponding classifier weight in the machine learning model (0063, 0067-68: weights for each layer are generated for forward propagation);
processing, iteratively, a plurality of frames of the sequential data using the machine learning model, wherein the machine learning model comprises the set of mask weights and the set of classifier weights, and wherein at each iteration the processing comprises (fig.4, 0067: iterative forward and back propagations steps):
generating a current frame of the plurality of frames of the sequential data (ibid: an input vector generated for the current sequential data frame);
determining a variance of classifier weights of the set of classifier weighs on features of the current frame (0063, 0065: determining current variance parameters), wherein the plurality of features includes the features of the current frame (ibid: weight mean and variance parameters for masks are determined for current frame being processed);
computing a penalty term over a data space of the current frame, wherein the of computing the penalty term over the data space of the current frame further comprises:
computing a penalty loss term over the data space of the current frame of the classifier weights across the current frame (0076-77: SNR function pushing for lower SNR constitutes a penalty loss term); and
computing a total loss as a function of a prediction error for mask weights of the set of mask weights on the features of the current frame and the penalty loss term (0074-75: combination with reconstruction loss as prediction error based on features); and
updating the mask weights using the classifier weights on the features and the penalty term (0073-74: mask weights are updated based on the penalty term / SNR loss during training), wherein:
the updating of the mask weights decreases the mask weights of spurious ones of the features according to the variance (0076-77: spurious features with greater variance are decreased, such as by reducing SNR by reducing mean parameter, hence, decreasing mask weight; see also 0080, 0089: reduction of mask weights via pruning) and increasing the mask weights of invariant ones of the features (ibid: certain invariant features are increased during training via backpropagation of the loss function, such as those leading to increase of the mean, decrease of variance during backpropagation of loss); and
outputting the machine learning model including the updated mask weights to a service, for performing a classification task based on a detection of at least one of the features in test data (fig.1, 0029 contemplates various applications of pruned inference system featuring the updated mask weights for performing classification tasks, these tasks requiring transforming inputs to intermediate features).
Lewis does not disclose: the invariant ones of the features include features whose target prediction probability is consistent across the plurality of environments, and
the spurious ones of the features include features whose prediction probabilities vary across the plurality of environments.
Kingma discloses: the invariant ones of the features include features whose target prediction probability is consistent across the plurality of environments (Lewis describes a total loss as a sum of a reconstruction loss corresponding to a basic prediction accuracy (eq.10, 0075) alongside a complexity loss (0076-78, eq.10-12), the complexity loss generally pushing the masking towards lower mean and higher variance regardless of the input data – see 0077, w_SNR = w_u^2 / w_sigma ^2. However, Lewis is not explicit on what kind of input will cause a decrease in this variance, i.e., how the accuracy loss will exert a counteracting push on the variance so that the variance doesn’t diverge.
Kingma §2-2.1 discloses estimating this log-likelihood term (eq.(2)), which corresponds to an accuracy loss (log p(y | x, w), see eq.2) via gradient descent (§2.1); in particular, eq.3 makes it clear that greater variance parameter, as encoded in phi, will lead to a greater loss, i.e., a greater departure from the labeled value y. That is, if a small variation on the weight (caused by the variance) would lead to greater divergence from the labeled training result y. Hence, Lewis discloses explicitly the diminishing effect that the expected log-likelihood loss has on the variance, which would otherwise be strictly pushed upwards or maximized by the KL divergence or the complexity loss of Lewis.
Hence, the loss function encourages lower variance on weights it considers critical to the calculation while the complexity loss encourages greater variance. Hence, the features and corresponding weights and masks considered to have a consistent or critical impact on the prediction across all environments will be increased (lower variance, not pruned, increasing mean according to the loss term), and
the spurious ones of the features include features whose prediction probabilities vary across the plurality of environments (Likewise, features that vary widely, and that have little influence on the prediction accuracy with high variance have their weights decreased (variance increased, mean diminished according to backpropagation, pruned)).
It would have been obvious before the effective filing date to one of ordinary skill in the art to modify the method of Lewis by incorporating the log-likelihood loss calculation of Lewis. Both concern the art of stochastic / Bayesian neural networks, and the incorporation would have, according to Lewis, allowing performing of stochastic gradient descent on phi, i.e., the mean and variance (§2.1).
For claim 2, Lewis modified by Kingma discloses the method of claim 1, as described above. Lewis further discloses: the data space of the current frame is an entire data space of the current frame (0027, 0061 contemplates training over entire space of training data);
the computer-implemented method includes no assumption on a degree of invariance of the mask weights (§fig.4, 0067: no assumption on invariance is made, relying instead on loss function 0076-77); and
the machine learning model is one of an existing model or a newly instantiated model (0034: initializing parameters for training a new model).
For claim 3, Lewis modified by Kingma discloses the method of claim 1, as described above. Lewis further discloses: wherein the computer-implemented method includes no labeled data about the plurality of environments (0061, 0067, 0076-77: no assumption about invariance or spuriousness is made as method relies on penalty loss for adjustment, no labels about environment type is incorporated).
For claim 9, Lewis modified by Kingma discloses the method of claim 1, as described above. Lewis further discloses: receiving the test data from at least one sensor (fig.1, 0028: sensor data) and
processing, using the machine learning model, the test data received from the at least one sensor (fig.1: sensor data is passed through inference model), wherein the machine learning model detects at least one event in the test data (0028-29: contemplates various event or event type predictions from sensor signals).
For claim 11, Lewis discloses: a computer readable medium comprising computer executable instructions which when executed by a computer system cause the computer system (fig.8, 0091-92) to perform a method for generating a machine learning model for classifying sequential data (fig.1, 0027-29 contemplates various data including sequential data such as text, video, audio, sensor, etc.), the method comprising:
accessing the sequential data, wherein the sequential data comprises a plurality of records having a plurality of features in a plurality of environments (ibid: sequential data encoded into vector form, hence, plurality of features; the cited passages also disclose a plurality of environments for each mode (e.g., different vehicle environments for automatic vehicle navigation, field data for forms, topics, etc.) as well as the modes themselves being different environments; see also 0061: training data from different scene environments);
initializing a set of mask weights on the plurality of features (0044-46: parameter tensors corresponding to divided volumes or sparsity constraints constitute pruning masks on the features / intermediate layer weights; see also 0063-66);
initializing a set of classifier weights on the plurality of features, wherein each mask weight of the set of mask weights corresponds to one of the classifier weight of the classifier weights, and each mask weight influences a corresponding classifier weight in the machine learning model (0063, 0067-68: weights for each layer are generated for forward propagation);
processing, iteratively, a plurality of frames of the sequential data using the machine learning model, wherein the machine learning model comprises the set of mask weights and the set of classifier weights, wherein at each iteration the processing comprises (fig.4, 0067: iterative forward and back propagations steps):
generating a current frame of the plurality of frames of the sequential data (ibid: an input vector generated for the current sequential data frame);
determining a mean and a variance of classifier weights of the set of classifier weights on the features of the current frame (0063, 0065: determining current mean and variance parameters), wherein the plurality of features includes the features of the current frame (ibid: weight mean and variance parameters for masks are determined for current frame being processed);
computing a penalty term over a data space of the current frame, wherein the computing of the penalty term over the data space of the current frame further comprises:
computing a penalty loss term over the data space of the current frame of the classifier weights across the current frame (0076-77: SNR function pushing for lower SNR constitutes a penalty loss term); and
computing a total loss as a function of a prediction error for mask weights of the set of mask weights on the features of the current frame and the penalty loss term (0074-75: combination with reconstruction loss as prediction error based on features); and
updating the mask weights using the classifier weights on the features and the penalty term (0073-74: mask weights are updated based on the penalty term / SNR loss during training), wherein:
the updating of the mask weights increases the mask weights of invariant ones of the features (0076-77: non-spurious weights have their mean and variance and scalar terms adjusted during the course of backpropagation, hence, portions of invariant ones will have increase mean and scalar weighting during backpropagation based on total loss function) and decreases the mask weights of spurious ones of the features according to the variance (0076-77: spurious features with greater variance are decreased, such as by reducing SNR by reducing mean parameter, hence, decreasing mask weight; see also 0080, 0089: reduction of mask weights via pruning); and
outputting the machine learning model including the updated mask weights to a service, for performing a classification task based on a detection of at least one of the features in test data (fig.1, 0029 contemplates various applications of pruned inference system featuring the updated mask weights for performing classification tasks, these tasks requiring transforming inputs to intermediate features).
Lewis does not disclose: the invariant ones of the features include features whose target prediction probability is consistent across the plurality of environments, and
the spurious ones of the features include features whose prediction probabilities vary across the plurality of environments.
Kingma discloses: the invariant ones of the features include features whose target prediction probability is consistent across the plurality of environments (Lewis describes a total loss as a sum of a reconstruction loss corresponding to a basic prediction accuracy (eq.10, 0075) alongside a complexity loss (0076-78, eq.10-12), the complexity loss generally pushing the masking towards lower mean and higher variance regardless of the input data – see 0077, w_SNR = w_u^2 / w_sigma ^2. However, Lewis is not explicit on what kind of input will cause a decrease in this variance, i.e., how the accuracy loss will exert a counteracting push on the variance so that the variance doesn’t diverge.
Kingma §2-2.1 discloses estimating this log-likelihood term (eq.(2)), which corresponds to an accuracy loss (log p(y | x, w), see eq.2) via gradient descent (§2.1); in particular, eq.3 makes it clear that greater variance parameter, as encoded in phi, will lead to a greater loss, i.e., a greater departure from the labeled value y. That is, if a small variation on the weight (caused by the variance) would lead to greater divergence from the labeled training result y. Hence, Lewis discloses explicitly the diminishing effect that the expected log-likelihood loss has on the variance, which would otherwise be strictly pushed upwards or maximized by the KL divergence or the complexity loss of Lewis.
Hence, the loss function encourages lower variance on weights it considers critical to the calculation while the complexity loss encourages greater variance. Hence, the features and corresponding weights and masks considered to have a consistent or critical impact on the prediction across all environments will be increased (lower variance, not pruned, increasing mean according to the loss term), and
the spurious ones of the features include features whose prediction probabilities vary across the plurality of environments (Likewise, features that vary widely, and that have little influence on the prediction accuracy with high variance have their weights decreased (variance increased, mean diminished according to backpropagation, pruned)).
It would have been obvious before the effective filing date to one of ordinary skill in the art to modify the method of Lewis by incorporating the log-likelihood loss calculation of Lewis. Both concern the art of stochastic / Bayesian neural networks, and the incorporation would have, according to Lewis, allowing performing of stochastic gradient descent on phi, i.e., the mean and variance (§2.1).
Claims 12-13, 19, 21 recite analogous subject matter to the above claims and hence are rejected under the same rationale.
Claim(s) 4, 14 are rejected under 35 U.S.C. 103 as being unpatentable over Lewis (US 20220237465 A1) in view of Kingma ("Variational dropout and the local reparameterization trick", published 2015) in view of Parascandolo ("Learning explanations that are hard to vary", published 2020).
For claim 4, Lewis modified by Kingma discloses the method of claim 1, as described above. Lewis modified by Kingma further does not disclose the limitations of claim 4.
Parascandolo discloses: wherein a number of iterations, s, is pre-defined for the processing of the plurality of frames (§3.3: §B.6 ¶3 discloses training for a preset number of epoch iterations, see also §3.3 disclosing the number of levels per epoch, sampling of a batch of data).
It would have been obvious before the effective filing date to one of ordinary skill in the art to modify method of Lewis modified by Kingma by incorporating the epoch training technique of Parascandolo. Both concern the art of neural network optimization, and the incorporation would have, according to Parascandolo, allow comparisons of learning rate for different hyperparameter sets (§3.3: §B.6 ¶3).
Claim(s) 14 recite analogous subject matter to the above claims and hence are rejected under the same rationale.
Claim(s) 7-8, 17-18 are rejected under 35 U.S.C. 103 as being unpatentable over Lewis (US 20220237465 A1) in view of Kingma ("Variational dropout and the local reparameterization trick", published 2015) in view of Cantarero (US 20160062967 A1).
For claim 7, Lewis modified by Kingma discloses the method of claim 1, as described above. Lewis modified by Kingma does not disclose the limitations of claim 7.
Cantarero discloses: wherein the test data includes a text corpus of documents (fig.1:102, 0033), and wherein the classification task comprises:
processing each of the documents using the machine learning model to identify instances of the features in each of the documents (fig.1:108-110, 0043-45: processing documents to identify features, such as genre, category, etc.; combination with Lewis yielding processing via the machine learning technique of Lewis);
classifying each of the documents according to the identified instances of the features in each of the documents (ibid: associating features with classifications such as sentiment); and
adding an indication of the classifying to each of the documents (fig.1:112, 0046).
It would have been obvious before the effective filing date to one of ordinary skill in the art to modify method of Lewis modified by Kingma by incorporating the sentiment analysis platform of Cantarero. Both concern the art of data science and machine analysis of human creations, and the incorporation would have, according to Cantarero, allow for fulfilment of the need for deeper analysis of textual data (0001-3).
For claim 8, Lewis modified by Kingma discloses the method of claim 1, as described above. Lewis modified by Kingma does not disclose the limitations of claim 8.
Cantarero discloses: wherein the test data includes a document (fig.1:102, 0033), and the classification task comprises:
processing the document using the machine learning model to identify instances of the features in the document (fig.1:108-110, 0043-45: processing documents to identify features, such as genre, category, etc.; combination with Lewis yielding processing via the machine learning technique of Lewis);
identifying a sentiment for each of a plurality of portions of the document according to corresponding ones of the identified instances of the features (ibid: identifying sentiment according to identified features; see also 0044: analyzing text into a plurality of portions); and
adding an indication of the sentiment for each portion of the plurality of portions to the document (fig.1:112, 0046: annotating of sentiment, see also 0044: transforming text stream into collection of information for categorization and anlysis, hence, adding sentiment corresponding to the analysis of the multiple portions).
It would have been obvious before the effective filing date to one of ordinary skill in the art to modify method of Lewis modified by Kingma by incorporating the sentiment analysis platform of Cantarero. Both concern the art of data science and machine analysis of human creations, and the incorporation would have, according to Cantarero, allow for fulfilment of the need for deeper analysis of textual data (0001-3).
Claim(s) 17-18 recite analogous subject matter to the above claims and hence are rejected under the same rationale.
Claim(s) 10, 20 are rejected under 35 U.S.C. 103 as being unpatentable over Lewis (US 20220237465 A1) in view of Kingma ("Variational dropout and the local reparameterization trick", published 2015) in view of Ma (US 20200192970 A1).
For claim 10, Lewis modified by Kingma discloses the method of claim 9, as described above. Lewis does not disclose the limitations of claim 10.
Ma discloses: wherein the at least one sensor includes an accelerometer and a gyroscope (35), and the at least one event is a selected from the group consisting of a walking event, a walking upstairs event, a walking downstairs event, a sitting event, a standing event, and a laying down event (0037).
It would have been obvious before the effective filing date to one of ordinary skill in the art to modify method of Lewis modified by Kingma by sensor-based detection technique of Ma. Both concern the art of machine pattern recognition, and the incorporation would have, according to Ma, allowed application to various human-worn devices to allow broad applicability such as human activity recognition (0033-34).
Claim(s) 20 recite analogous subject matter to the above claims and hence are rejected under the same rationale.
Response to Arguments
Applicant’s arguments have been fully considered. In the remarks, Applicant argues:
1) The amended claims overcomes the art of record.
Applicant’s arguments are moot in view of newly cited art.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Chattopadhyay ("Learning to balance specificity and invariance for in and out of domain generalization", published 2020) discloses a technique of learning domain-specific masks.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to LIANG LI whose telephone number is (303)297-4263. The examiner can normally be reached Mon-Fri 9-12p, 3-11p MT (11-2p, 5-1a ET).
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 extension fee 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 date of this final action.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor Jennifer Welch can be reached on (571)272-7212. 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 Patent Center and the Private Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from Patent Center or Private PAIR. Status information for unpublished applications is available through Patent Center or Private PAIR to authorized users only. Should you have questions about access to Patent Center or the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free).
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. The examiner is available for interviews Mon-Fri 6-11a, 2-7p MT (8-1p, 4-9p ET).
/LIANG LI/
Primary examiner AU 2143
