DETAILED ACTION
This Office Action is sent in response to the Applicant’s Communication received on 02/03/2026 for application number 18/070,923. The Office hereby acknowledges receipt of the following and placed of record in file: Specification, Drawings, Abstract, Oath/Declaration, IDS, and Claims.
Claim 17-20 are canceled.
Claims 1, 2, 5-7, 11, 14-16 are amended.
Claims 1-16 are pending.
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 .
Priority
Copies of all priority documents have been received.
Response to Arguments
35 USC 101
On page 11 of the remarks section, the Applicant argues that the claimed embodiments clearly realize improvements to the technology of training an artificial neural network, such as the robustness of the neural network at the end of its learning
phase (paras. [0001-[0004]) by performing a secondary training to detect weaknesses in the model and to reinforce the model using targeted data and an adapted learning rate by increasing the learning rate of the output neurons of the network that are associated with property values that are most often misestimated (paras. (0023]-[0024]). Independent claims 1, 14 and 16 explicitly recite process steps that are performed to realize the technological improvement. For example, in claim 1, after implementing a primary training of a neural network to identify at least one target property value from a first set of primary training data labeled by associating these data with a first set of target property values, the claimed process implements secondary training that includes: obtaining, for at least a first target value of a target property, a set of secondary training data; identifying an estimated value of the target property associated with at least one said secondary training data, by using the artificial neural network trained by said primary training; for each estimated value different from said first target value, obtaining a number of confusions corresponding to a number of times said estimated value has been estimated; and increasing learning rates associated with said neurons of the output layer of said neural network that are associated with values of interest corresponding to estimated values with largest numbers of confusions.
On page 12 of the remarks section, the Applicant further argues that, regarding the embodiments of claims 2 and 15, the artificial neural network is further configured to identify a digital use among a set of digital uses, each digital use being described by a digital behavior associated with at least one said property and by a digital environment associated with at least one said property (paras. [0027]-[0047]). The primary training identifies properties describing the digital use including at least one property describing a digital behavior and a digital environment of this digital use (para. [0036]). Examples of the digital use include a used application category, a used application, an operation implemented at the application level, a user interaction state, a used device, a used operating system, a used browser, and at least one related characteristic of a communication network, for example an Internet connection speed or a latency (paras. [0038]-[0046]). The claimed process improves the reliability and the accuracy of the identification of the digital use (paras. [0037] and [0047]). Furthermore, the claimed process allows for the reduction of the trained neural network, which reduces the load on computational resources (para. [0037]).
The Examiner finds the Applicant’s argument persuasive. Therefore, the 35 USC 101 rejections is withdrawn.
35 USC 112
The Applicant argues that the plain meaning of the feature indicates that the underlined "that are associated with" refers to the neurons of the output layer rather than the learning rates, particularly when read in light of the specification. For example, para. [0024] of the published application (US20230169338) explains that it is the neurons that are associated with property values that are most often misestimated (i.e., largest numbers of confusions) that have their associated learning rates increased.
The Examiner respectfully disagrees. As mention in the rejection dated 10/06/2025, it is unclear if the limitation “that are associated with” is referring to “the learning rates” or “neurons of the output layer of said neural network.” The Applicant cited paragraph 0024 of published application US20230169338 does not cure the deficiencies of the current claim indefiniteness. Specifically, it can be seen from the table below that the claim limitation at issue relates to “associated with values of interest corresponding to estimated values with largest numbers of confusions”, while Applicant cited paragraph 0024 of published application US20230169338 relates to “associated with property values most often misestimated”.
Claim 1
Paragraph 0024 of published application US20230169338
increasing learning rates associated with said neurons of the output layer of said neural network that are associated with values of interest corresponding to estimated values with largest numbers of confusions
This reinforcement is performed by increasing the learning rate of the output neurons of the network that are associated with property values most often misestimated
Therefore, the description at paragraph 0024 does not help overcome the 35 USC 112B indefiniteness rejection.
35 USC 103
On page 15 of the remarks section, Applicant states that one having ordinary skill in the art would not have been motivated to combine the multiple references to form the claimed embodiments because they address fundamentally different problems, operate at different stages of the machine-learning lifecycle, and rely on distinct technical mechanisms. For example, Packes is concerned with system-level training, retraining, and deployment of neural networks for real-estate valuation; Aceto is directed to post-training performance evaluation and calibration analysis; Mittal addresses input-space data augmentation for long-tailed datasets using frozen models; and Popescu discloses a conventional, time-based global learning-rate schedule.
The Examiner respectfully disagrees. The test for obviousness is not whether the features of a secondary reference may be bodily incorporated into the structure of the primary reference; nor is it that the claimed invention must be expressly suggested in any one or all of the references. Rather, the test is what the combined teachings of the references would have suggested to those of ordinary skill in the art. See In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981). Specifically, Examiner is combining the methodology of for estimated value different from target value, obtaining a number of confusions corresponding to a number of times estimated value has been estimated of Aceto, learning rates associated with values of interest correspond to estimated values with largest numbers of confusions of Mittal, and increasing the learning rates of Popescu with the teaching of Packes. Examiner is not proposing bodily incorporation of Aceto-Mittal-Popescu to Packes.
The Applicant further argues that Packes does not disclose learning rates at all, nor does it disclose any association between learning rates and individual neurons, and certainly not output-layer neurons associated with specific estimated values. The passages cited by the Examiner, including paragraph 0071, merely disclose that a
new training process may be started or that sets of neural networks may be updated when changes in data reach a threshold. Restarting training, retraining networks, updating weights, or switching training methods are network-level or process-level operations. This disclosure of Packes neither inherently nor explicitly teach learning rates, and they do not teach learning rates that are associated with neurons of the output layer. The Examiner's reasoning improperly equates the existence of training or retraining with the existence of neuron-associated learning rates, which is an unsupported inference. Training can be performed using many mechanisms that do not involve learning rates associated with individual neurons, and Packes does not disclose such an association. Moreover, Packes does not disclose any functional linkage between estimated values produced by the neural network and learning behavior at the level of output neurons. In Packes, estimated values are used for evaluation, reporting, aggregation, prediction, or triggering retraining processes, but not for directly influencing learning behavior associated with specific output neurons. The claimed method, by contrast, requires that estimated values obtained by applying the trained neural network to secondary training data be used to act on learning rates associated with output-layer neurons corresponding to values of interest. This structural and functional linkage is absent from Packes.
The Examiner respectfully disagrees. First, the plain meaning of the term “learning rates” refers to shifting model parameters in order to improve the model, so Packes does indeed teach the limitation “learning rates that are associated with neurons of the output layer of neural network that are associated with values of interest corresponding to estimated values” in paragraph 0071, “An agent may periodically verify the changes rate (e.g., percentage of the number of records changed from the entire dataset) in the properties information (values of interest) used for the training. If this rate hits a certain level, a new process for network training may be started (learning rates) and (associated with) a set of neuronal networks (neurons… of neural network) may then be updated…The two sets of neuronal networks (e.g., any RVE neural networks (e.g., as described with respect to process 400) that may be used for obtaining an estimated property value (output layer of neural network… corresponding to estimated values) (e.g., at step 505), which may be feed forward networks for estimation”. Second, while a “functional linkage between estimated values produced by the neural network and learning behavior at the level of output neurons”, “directly influencing learning behavior associated with specific output neurons”, “secondary training data be used to act on learning rates associated with output-layer neurons corresponding to values of interest” and other “structural and functional linkages” appear to be disclosed invention, they are not explicitly recited in the claimed limitation. In response to applicant's argument that the references fail to show certain features of the invention, it is noted that the features upon which applicant relies are not recited in the rejected claim(s). Although the claims are interpreted in light of the specification, limitations from the specification are not read into the claims. See In re Van Geuns, 988 F.2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993).
On page 17 of the remarks section, the Applicant further argues that reference Aceto’s teaching of claim 1’s limitations does not involve, describe, or require computing class-to-class confusions or counting how many times a particular estimated value is produced instead of a target value. The Examiner's reliance on this paragraph to support the notion of "obtaining a number of confusions corresponding to a number of times an estimated value has been estimated" is technically incorrect. While the Examiner correctly recognizes that Aceto does not teach increasing learning rates based on confusions, Aceto also does not disclose the confusion-counting concept itself and therefore cannot be used to supply that feature when combined with Packes.
The Examiner respectfully disagrees. First, the term “confusions” is broad, and under the broadest reasonable interpretation, Aceto does indeed teach “obtaining a number of confusions corresponding to a number of times an estimated value has been estimated” in section IV (B), paragraph 4: “To deepen the performance evaluation, Fig. 2 reports a calibration analysis, assessing (obtaining) whether the class-probability estimates are representative of the true-class (posterior) probabilities, where a miscalibrated classifier (a number of confusions) (on a given task) returns excessively optimistic or pessimistic decisions (corresponding to a number of times an estimated value has been estimated)”. Second, although “class-to-class confusions or counting how many times a particular estimated value is produced instead of a target value” appears to be disclosed invention, it is not explicitly recited in the claimed limitation. In response to applicant's argument that the references fail to show certain features of the invention, it is noted that the features upon which applicant relies are not recited in the rejected claim(s). Although the claims are interpreted in light of the specification, limitations from the specification are not read into the claims. See In re Van Geuns, 988 F.2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993).
On page 18 of the remarks section, Applicant further argues that The Examiner's interpretation of Mittal's "learning rate" is technically incorrect. In Mittal, the parameter
δ
is not a training learning rate of the neural network, but a fixed step size used inside a gradient-ascent image perturbation procedure. This parameter controls how strongly the input image is modified during the hallucination process, while the network weights are explicitly frozen. It is therefore neither associated with neurons of the output layer, nor associated with class values, nor used to control the learning dynamics of the model during training.
The Examiner respectfully disagrees. Although appears to be disclosed invention, the claimed invention does not explicitly recite the limitations “a training learning rate of the neural network”, a learning rate “associated with class values”, a learning rate “used to control the learning dynamics of the model during training”. In response to applicant's argument that the references fail to show certain features of the invention, it is noted that the features upon which applicant relies are not recited in the rejected claim(s). Although the claims are interpreted in light of the specification, limitations from the specification are not read into the claims. See In re Van Geuns, 988 F.2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993). Moreover, Mittal alone was not cited to teach the limitation “increasing learning rates associated with said neurons of the output layer of said neural network that are associated with values of interest corresponding to estimated values with largest numbers of confusions”. Rather, the combination Aceto-Mittal-Popescu teach the aforementioned limitation. Aceto teaches “for estimated value different from target value, obtaining a number of confusions corresponding to a number of times estimated value has been estimated” in section IV (B), paragraph 4: “To deepen the performance evaluation, Fig. 2 reports a calibration analysis, assessing (obtaining) whether the class-probability estimates are representative of the true-class (posterior) probabilities, where a miscalibrated classifier (a number of confusions) (on a given task) returns excessively optimistic or pessimistic decisions (corresponding to a number of times an estimated value has been estimated)”. Mittal teaches “learning rates associated with values of interest correspond to estimated values with largest numbers of confusions” in paragraph 0048: “The disclosed implementation of BLT selects a confusing class in step 4 by using information from the confusion matrix C (associated with values of interest) for a given tail class c… This strategy will select the most confusing classes (estimated values with largest numbers of confusions) more often… The disclosed implementation runs the gradient ascent image generation procedure with a learning rate δ=0.7 (learning rates). It stops running when S.sub.c′(I′)≥s.sub.c′ or when it reaches 15 iterations”. Popescu teaches “increasing the learning rates” in paragraph 0172: “During the first five epochs of training, the learning rate will increase linearly from 0 to 4×10.sup.−3”. 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).
The Applicant further argues that Mittal does not associate learning rates with values of interest corresponding to estimated values having the largest numbers of confusions. The confusion matrix in Mittal is used only to bias the selection of which confusing class is targeted more often during image generation. Once that class is selected, the same parameter
δ
is used, independently of how confusing that class is, and independently of any output neuron. There is no disclosure in Mittal of adapting, assigning, or tuning learning rates based on confusion counts or on specific output
values.
The Examiner respectfully disagrees. Mittal indeed teaches “learning rates associated with values of interest correspond to estimated values with largest numbers of confusions” in paragraph 0048: “The disclosed implementation of BLT selects a confusing class in step 4 by using information from the confusion matrix C (associated with values of interest) for a given tail class c… This strategy will select the most confusing classes (estimated values with largest numbers of confusions) more often… The disclosed implementation runs the gradient ascent image generation procedure with a learning rate δ=0.7 (learning rates). It stops running when S.sub.c′(I′)≥s.sub.c′ or when it reaches 15 iterations”. Moreover, although appears to be disclosed invention, the limitations “adapting, assigning, or tuning learning rates based on confusion counts or on specific output values” is not explicitly recited in the claim. In response to applicant's argument that the references fail to show certain features of the invention, it is noted that the features upon which applicant relies are not recited in the rejected claim(s). Although the claims are interpreted in light of the specification, limitations from the specification are not read into the claims. See In re Van Geuns, 988 F.2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993).
The Applicant further argues that Mittal cannot remedy the deficiencies of Packes and/or Aceto in the manner suggested by the Examiner. Mittal operates in a fundamentally different technical loop, focused on input-space data augmentation with frozen model parameters, whereas Packes concerns retraining a model and Aceto concerns performance evaluation via calibration. Combining these documents would require reinterpreting Mittal's image perturbation step size as a model learning rate and relocating it from input generation to output-layer training, which is neither taught nor suggested in Mittal.
The Examiner respectfully disagrees. As mentioned above, the test for obviousness is not whether the features of a secondary reference may be bodily incorporated into the structure of the primary reference; nor is it that the claimed invention must be expressly suggested in any one or all of the references. Rather, the test is what the combined teachings of the references would have suggested to those of ordinary skill in the art. See In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981). Specifically, as mentioned above, Examiner is combining the methodology of, for estimated value different from target value, obtaining a number of confusions corresponding to a number of times estimated value has been estimated of Aceto, learning rates associated with values of interest correspond to estimated values with largest numbers of confusions of Mittal, and increasing the learning rates of Popescu to the teaching of Packes. Examiner is not proposing bodily incorporation of Aceto-Mittal-Popescu to Packes. The methodology of Aceto would deepen the performance evaluation of Packes’ system/method, the methodology of Mittal would provide a means for learning from misclassification in order to improve model results of Packes’ system/method, and the methodology of Popescu would provide a means for learning from misclassification in order to improve model results of Packes’ system/method. The combined teachings of the references Packes-Aceto-Mittal-Popescu, therefore would have suggested to those of ordinary skill in the art.
On page 19 of the remarks section, The Applicant further argues that Popescu’s learning-rate adjustment is purely time-based and applies uniformly to the entire model during training, without any dependence on predicted values, output neurons, confusion patterns, or class specific behavior. Popescu does not disclose selecting learning rates based on estimated values, values of interest, or misclassification frequency, nor does it associate learning-rate increases with particular outputs or properties identified by the network. Accordingly, Popescu teaches only a standard, global optimization heuristic for stabilizing training and improving convergence, and does not provide any mechanism for increasing learning rates in response to confusion or for specific values of interest, as required by the claimed invention.
The Examiner respectfully disagrees. Popescu alone was not cited as teaching the limitation “increasing learning rates associated with said neurons of the output layer of said neural network that are associated with values of interest corresponding to estimated values with largest numbers of confusions”. Rather, as mentioned above, the combination Aceto-Mittal-Popescu teach the aforementioned limitation. Aceto teaches “for estimated value different from target value, obtaining a number of confusions corresponding to a number of times estimated value has been estimated” in section IV (B), paragraph 4: “To deepen the performance evaluation, Fig. 2 reports a calibration analysis, assessing (obtaining) whether the class-probability estimates are representative of the true-class (posterior) probabilities, where a miscalibrated classifier (a number of confusions) (on a given task) returns excessively optimistic or pessimistic decisions (corresponding to a number of times an estimated value has been estimated)”. Mittal teaches “learning rates associated with values of interest correspond to estimated values with largest numbers of confusions” in paragraph 0048: “The disclosed implementation of BLT selects a confusing class in step 4 by using information from the confusion matrix C (associated with values of interest) for a given tail class c… This strategy will select the most confusing classes (estimated values with largest numbers of confusions) more often… The disclosed implementation runs the gradient ascent image generation procedure with a learning rate δ=0.7 (learning rates). It stops running when S.sub.c′(I′)≥s.sub.c′ or when it reaches 15 iterations”. Popescu teaches “increasing the learning rates” in paragraph 0172: “During the first five epochs of training, the learning rate will increase linearly from 0 to 4×10.sup.−3”. 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). Moreover, although appears to be disclosed invention, limitations such as “selecting learning rates based on estimated values, values of interest, or misclassification frequency”, “learning-rate increases with particular outputs or properties identified by the network”, and “mechanism for increasing learning rates in response to confusion or for specific values of interest” are explicitly recited in the claimed limitation. In response to applicant's argument that the references fail to show certain features of the invention, it is noted that the features upon which applicant relies are not recited in the rejected claim(s). Although the claims are interpreted in light of the specification, limitations from the specification are not read into the claims. See In re Van Geuns, 988 F.2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993).
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 1-16 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Example claim 1 states in part, “…increasing the learning rates associated with said neurons of the output layer of said neural network that are associated with values of interest corresponding to estimated values with the largest numbers of confusions.” It is unclear if the limitation “that are associated with” is referring to “the learning rates” or “neurons of the output layer of said neural network.” Therefore, the scope and the metes and bounds of the claim are not ascertainable. As such, the claim is indefinite. For the purpose of the examination, the examiner interprets the limitations to be “increasing the learning rates associated with said neurons of the output layer of said neural network that (interpreted as learning rates) are associated with values of interest corresponding to estimated values with the largest numbers of confusions.”
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claim(s) 1, 2, 5-10, and 13-16 are rejected under 35 U.S.C. 103 as being unpatentable over Packes et al. (US 20150242747 A1), hereinafter Packes, in view of Aceto et al. (Encrypted Multitask Traffic Classification via Multimodal Deep Learning, published August 6, 2021), hereinafter Aceto, Mittal et al. (US 20220414392 A1), hereinafter Mittal, and Popescu et al. (US 20220383986 A1), hereinafter Popescu.
Regarding claim 1, Packes teaches,
A computer-implemented training method for training an artificial neural network so that said artificial neural network identifies at least one value of a property among a plurality of property values, each property being able to take at least two different values, said artificial neural network including an output layer including, for at least one said property value, a neuron configured to deliver a prediction score for said property value [Para 0020, Para 0020, the REP may train, retrain, and utilize sets of neural networks to estimate values of real estate properties and/or to predict values of real estate properties into the future… A neural network or neuronal network or artificial neural network may be hardware-based, software-based, or any combination thereof, such as any suitable model (e.g., a computational model), which, in some embodiments, may include one or more sets or matrices of weights; Para 0035, historical data may lead to better results when training and/or retraining a neural network to estimate property value based on differences in attribute values. In one embodiment, to prepare historical data, historical property values (e.g., historical prices) may be analyzed to determine property values as of a specified (e.g., current) time period. In one implementation, historical data may be prepared using the following preparation process. An estimation time period unit (e.g., one month) may be defined. Historical data may be obtained for a specified estimation time frame (e.g., the last year) for properties that have data regarding property values during the estimation time frame (e.g., properties that were sold during the last year and have a selling price, properties whose property values were evaluated during a previous preparation iteration, etc.). The obtained historical data may be sliced for each estimation time period (e.g., for each month during the last year); Para 0040, the neural network parameters may include the number of neurons in the output layer],
said method implemented by a training system [Para 0020, the REP may train, retrain, and utilize sets of neural networks to estimate values of real estate properties and/or to predict values of real estate properties into the future] and comprising: implementing a primary training comprising training said neural network to identify at least one target property value from a first set of primary training data labeled by associating these data with a first set of target property values [Para 0035, before it is used (e.g., as inputs, as outputs, etc.), historical data may be prepared to reduce property value variation associated with time. For example, reducing property value variation associated with time (e.g., inflation, housing market trends, etc.) in historical data may lead to better results when training and/or retraining a neural network to estimate property value based on differences in attribute values]; and
implementing a secondary training including: obtaining, for at least a first target value of a target property, a set of secondary training data; identifying an estimated value of the target property associated with at least one said secondary training data, by using the artificial neural network trained by said primary training [Para 0058, Attribute values for the property may be augmented at step 405 of process 400… if the user modifies an attribute value, the modified attribute value may replace the attribute value stored in the data store; Para 0020, the REP may train, retrain, and utilize sets of neural networks to estimate values of real estate properties and/or to predict values of real estate properties into the future… A neural network or neuronal network or artificial neural network may be hardware-based, software-based, or any combination thereof, such as any suitable model (e.g., a computational model), which, in some embodiments, may include one or more sets or matrices of weights];
learning rates (Para 0071, a new process for network training may be started) associated with neurons of the output layer of neural network that are associated with values of interest corresponding to estimated values [Para 0071, An agent may periodically verify the changes rate (e.g., percentage of the number of records changed from the entire dataset) in the properties information used for the training. If this rate hits a certain level, a new process for network training may be started and a set of neuronal networks may then be updated. The system may be scheduled to execute periodical analysis. This analysis can reflect future price increasing for certain types of properties or can reveal trends in the real estate market. When these results are available, the system can notify its clients or can push the results to the registered clients. The two sets of neuronal networks (e.g., any RVE neural networks (e.g., as described with respect to process 400) that may be used for obtaining an estimated property value (e.g., at step 505), which may be feed forward networks for estimation].
Packes teaches the above limitations of claim 1, but does not teach for each estimated value different from said first target value, obtaining a number of confusions corresponding to a number of times said estimated value has been estimated; and Increasing learning rates associated with values of interest correspond to estimated values with largest numbers of confusions.
Aceto teaches,
for estimated value different from target value, obtaining a number of confusions corresponding to a number of times estimated value has been estimated [Sect IV (B), para 4, To deepen the performance evaluation, Fig. 2 reports a calibration analysis, assessing whether the class-probability estimates are representative of the true-class (posterior) probabilities, where a miscalibrated classifier (on a given task) returns excessively optimistic or pessimistic decisions.];
Aceto is analogous to the claimed invention as they both relate to deep learning. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Packes’ teachings to incorporate the teachings of Aceto and provide a confusion matrix [Aceto, Sect IV (B), para 4] in order to deepen the performance evaluation.
Packes-Aceto do not teach Increasing learning rates associated with values of interest correspond to estimated values with largest numbers of confusions.
Mittal teaches,
learning rates associated with values of interest correspond to estimated values with largest numbers of confusions [Para 0048, The disclosed implementation of BLT selects a confusing class in step 4 by using information from the confusion matrix C for a given tail class c… This strategy will select the most confusing classes more often… The disclosed implementation runs the gradient ascent image generation procedure with a learning rate δ=0.7. It stops running when S.sub.c′(I′)≥s.sub.c′ or when it reaches 15 iterations].
Mittal is analogous to the claimed invention as they both relate to identifying confusions. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Packes and Aceto’s teachings to incorporate the teachings of Mittal and provide learning rates associated with values of interest correspond to estimated values with largest numbers of confusions in order to provide a means for learning from misclassification in order to improve model results.
Packes-Aceto-Mittal do not teach increasing the learning rates.
Popescu teaches,
increasing the learning rates [Para 0172, During the first five epochs of training, the learning rate will increase linearly from 0 to 4×10.sup.−3].
Popescu is analogous to the claimed invention as they both relate to utilizing learning rates for training. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Packes, Aceto, and Mittal’s teachings to incorporate the teachings of Popescu and provide increasing learning rates in order to learn from desirable weights.
Regarding claim 2, Packes-Aceto-Mittal-Popescu teach the limitations of claim 1.
Packes further teaches,
wherein said artificial neural network is configured to identify a digital use among a set of digital uses, each digital use being described by a digital behavior associated with at least one said property and by a digital environment associated with at least one said property [Para 0057, a user may utilize a website, a mobile app, an external application, and/or the like to specify any suitable attribute values for any suitable set of attributes. In one implementation, the user may specify attribute values for any of the attributes discussed with regard to step 105 of process 100. For example, the user may be enabled to enter attribute values for a minimum number of attributes (e.g., one or any other suitable number greater than one). In another example, the user may enter attribute values for a greater number of attributes to enhance the accuracy of the price prediction. The REP may be operative to enable a user to enter a minimum amount of information to allow the REP to identify a property (e.g., the address)];
said primary training comprising training said neural network to identify at least one target digital use among said set of digital uses, the data of said first set of primary training data being extracted from a first plurality of network packets captured during at least one execution of an application associated with said first target digital use, said primary training data being labeled by associating these data with a first set of target property values describing at least a digital behavior and a digital environment of said first target digital use; and the data of said set of secondary training data obtained for said at least one first target value being extracted from a second plurality of network packets captured during at least one execution of an application associated with a digital use whose said target property is described by said first target value [Para 0035, historical data may be prepared to reduce property value variation associated with time. For example, reducing property value variation associated with time (e.g., inflation, housing market trends, etc.) in historical data may lead to better results when training and/or retraining a neural network to estimate property value based on differences in attribute values. In one embodiment, to prepare historical data, historical property values (e.g., historical prices) may be analyzed to determine property values as of a specified (e.g., current) time period; Para 0040, Neural network parameters may be obtained at step 117 of process 100. In one implementation, the neural network parameters may be obtained from the administrator or any other suitable party via any suitable GUI of the REP. In another implementation, the neural network parameters may be obtained from a configuration file. All the parameters may be saved in a database. When a network is trained, all the information about this process may be saved in the database (e.g., the training parameters, the data set used, the network performances, training execution time, all the testing results for that specific network with results for each record, etc.); Para 0057, a user may utilize a website, a mobile app, an external application, and/or the like to specify any suitable attribute values for any suitable set of attributes; Para 0058, Attribute values for the property may be augmented at step 405 of process 400; Para 0094, the input/output devices may include one or more network devices 815… In one implementation, a network device may be a network card that may obtain (e.g., via a Category 5 Ethernet cable), process, output (e.g., via a wireless antenna), and/or the like network data (e.g., REP data)… A network card may be a… wireless modem based on cellular protocols such as high speed packet access (HSPA)].
Regarding claim 5, Packes-Aceto-Mittal-Popescu teach the limitations of claim 1.
Packes further teaches,
wherein said primary training includes: obtaining said first set of primary training data [Para 0035, Historical data may be obtained for a specified estimation time frame (e.g., the last year) for properties that have data regarding property values during the estimation time frame];
applying said first set of primary training data to an input of the artificial neural network; and modifying at least one weight of the artificial neural network as a function of the first set of target property values and of the prediction scores obtained [Para 0035, prepared historical data with property values as of the specified (e.g., current) time period may be used (e.g., as inputs, as outputs) when training and/or retraining the set of neural networks; Para 0020, A neural network or neuronal network or artificial neural network may be hardware-based, software-based, or any combination thereof, such as any suitable model (e.g., a computational model), which, in some embodiments, may include one or more sets or matrices of weights; Para 0039, the obtained training method parameters may include the number of epochs to use (e.g., the number of cycles the algorithm will work on trying to minimize the output error by changing the weights matrix].
Regarding claim 6, Packes-Aceto-Mittal-Popescu teach the limitations of claim 1.
Packes further teaches,
wherein the properties defining the digital use comprise at least one property among the following properties:
a used application category,
a used application,
an operation implemented at an application level,
a user interaction state,
a used device,
a used operating system,
a used browser, and
at least one related characteristic of a communication network
[Para 0057, a user may utilize a website, a mobile app, an external application, and/or the like to specify any suitable attribute values for any suitable set of attributes].
Regarding claim 7, Packes-Aceto-Mittal-Popescu teach the limitations of claim 1.
Packes further teaches,
wherein said first sets of primary and secondary training data comprise, for each packet of the first and second plurality of network packets, at least one training data among the following training data:
the size of the packet,
a duration between receiving or sending the packet and receiving or sending a previous packet of a same session or of a same protocol as said packet [Para 0035, before it is used (e.g., as inputs, as outputs, etc.), historical data may be prepared to reduce property value variation associated with time. For example, reducing property value variation associated with time (e.g., inflation, housing market trends, etc.) in historical data may lead to better results when training and/or retraining a neural network to estimate property value based on differences in attribute values. In one embodiment, to prepare historical data, historical property values (e.g., historical prices) may be analyzed to determine property values as of a specified (e.g., current) time period… An estimation time period unit (e.g., one month) may be defined. Historical data may be obtained for a specified estimation time frame (e.g., the last year) for properties that have data regarding property values during the estimation time frame (e.g., properties that were sold during the last year and have a selling price, properties whose property values were evaluated during a previous preparation iteration, etc.).],
a source port of the packet,
a destination port of the packet,
a direction of the packet [Para 0034, historical data may be collected (e.g., continuously, periodically, etc.) from a plurality of sources such as an Automated City Register Information System (ACRIS), the Department of Buildings of the City of New York, Building Owners' and Brokers' web sites, the New York State Attorney General's Office, the Public Library, Federal Energy Management Program, real estate news publications, and/or the like], and
a protocol of a higher-level layer in the packet.
Regarding claim 8, Packes-Aceto-Mittal-Popescu teach the limitations of claim 1.
Packes further teaches,
wherein said primary training is reiterated for at least a second set of primary training data associated with a second set of target property values describing a second digital use [Para 0020, the REP may train, retrain, and utilize sets of neural networks to estimate values of real estate properties and/or to predict values of real estate properties into the future].
Regarding claim 9, Packes-Aceto-Mittal-Popescu teach the limitations of claim 1.
Packes further teaches,
wherein said secondary training is reiterated for at least a second target value of a target property [Para 0020, the REP may train, retrain, and utilize sets of neural networks to estimate values of real estate properties and/or to predict values of real estate properties into the future].
Regarding claim 10, Packes-Aceto-Mittal-Popescu teach the limitations of claim 1 including the primary training (Packes: para 0035).
Aceto further teaches,
wherein training uses a multitask learning [Sect II, last para, we adopt a multimodal architecture to process input data—the combination of (i) and (ii) allows for automatic feature learning from heterogeneous data input; (iii) we feed the architecture with simple unbiased inputs; Sect III (A), para 1, we describe the proposed methodology for multipurpose encrypted TC via multimodal multitask DL].
Aceto is analogous to the claimed invention as they both relate to deep learning. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Packes’ teachings to incorporate the teachings of Aceto and provide multitask learning [Aceto, Abstract] to improve model analyzation by learning both intra- and inter-modality dependencies simultaneously.
Regarding claim 13, Packes-Aceto-Mittal-Popescu teach the limitations of claim 2.
Aceto further teaches,
filtering at least one network packet which is not associated with an operation implemented by a user at the application level [Sect III (A), para 2, As a necessary prerequisite, raw traffic is first segmented via traffic object segmentation into distinct TC objects [1], each constituting a subset of network packets sharing some common properties; Sect IV (A), para 2, we have performed a cleaning operation to remove this noisy traffic and make our results more meaningful].
Aceto is analogous to the claimed invention as they both relate to deep learning. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Packes’ teachings to incorporate the teachings of Aceto and provide filtering network packets [Aceto, Para IV (A), para 2] in order to make results more meaningful.
Regarding claim 14, Packes further teaches,
A training system for training an artificial neural network [Para 0025, the method includes training, with the system, the neural network] so that said artificial neural network identifies at least one property value among a plurality of property values, each property being able to take at least two different values, the artificial neural network including an output layer including, for at least one said property value , a neuron configured to deliver a prediction score for said property value [Para 0020, Para 0020, the REP may train, retrain, and utilize sets of neural networks to estimate values of real estate properties and/or to predict values of real estate properties into the future… A neural network or neuronal network or artificial neural network may be hardware-based, software-based, or any combination thereof, such as any suitable model (e.g., a computational model), which, in some embodiments, may include one or more sets or matrices of weights; Para 0035, historical data may lead to better results when training and/or retraining a neural network to estimate property value based on differences in attribute values. In one embodiment, to prepare historical data, historical property values (e.g., historical prices) may be analyzed to determine property values as of a specified (e.g., current) time period. In one implementation, historical data may be prepared using the following preparation process. An estimation time period unit (e.g., one month) may be defined. Historical data may be obtained for a specified estimation time frame (e.g., the last year) for properties that have data regarding property values during the estimation time frame (e.g., properties that were sold during the last year and have a selling price, properties whose property values were evaluated during a previous preparation iteration, etc.). The obtained historical data may be sliced for each estimation time period (e.g., for each month during the last year); Para 0040, the neural network parameters may include the number of neurons in the output layer],
said system comprising: at least one processor [Para 0022, a processor]; and at least one non-transitory computer readable medium comprising instructions stored thereon [Para 0026, a non-transitory computer-readable medium may include computer-readable instructions recorded thereon] which when executed by the at least one processor configure the system to implement a method comprising:
obtaining a first set of primary training data labeled by associating these data with a first set of target property values [Para 0035, before it is used (e.g., as inputs, as outputs, etc.), historical data may be prepared to reduce property value variation associated with time. For example, reducing property value variation associated with time (e.g., inflation, housing market trends, etc.) in historical data may lead to better results when training and/or retraining a neural network to estimate property value based on differences in attribute values];
applying the first set of primary training data to an input of the artificial neural network, to train, during a primary training, said neural network to identify at least one said target property value; modifying at least one weight of the artificial neural network as a function of the first set of target property values and of the prediction scores obtained [Para 0035, prepared historical data with property values as of the specified (e.g., current) time period may be used (e.g., as inputs, as outputs) when training and/or retraining the set of neural networks; Para 0020, A neural network or neuronal network or artificial neural network may be hardware-based, software-based, or any combination thereof, such as any suitable model (e.g., a computational model), which, in some embodiments, may include one or more sets or matrices of weights; Para 0039, the obtained training method parameters may include the number of epochs to use (e.g., the number of cycles the algorithm will work on trying to minimize the output error by changing the weights matrix];
obtaining, for at least a first target value of a target property, a set of secondary training data [Para 0058, Attribute values for the property may be augmented at step 405 of process 400… if the user modifies an attribute value, the modified attribute value may replace the attribute value stored in the data store; Para 0020, the REP may train, retrain, and utilize sets of neural networks to estimate values of real estate properties and/or to predict values of real estate properties into the future];
applying, during a secondary training, at least one data of the set of secondary training data to the input of the neural network to identify an estimated value of said target property associated with at least one said secondary training data, by using the artificial neural network trained by said primary training [Para 0058, Attribute values for the property may be augmented at step 405 of process 400. In one implementation, default values (e.g., typical attribute values for similar properties) may be substituted for attribute values not provided by the user. In one embodiment, the user may enter attribute values for a property recognized (e.g., based on the address) by the REP (e.g., information regarding the property may be stored in the data sets data store 830d). Accordingly, the REP may retrieve such stored attribute value information and populate relevant fields (e.g., of any suitable GUI) with the retrieved attribute values]; and
learning rates (Para 0071, a new process for network training may be started) associated with neurons of the output layer of neural network that are associated with values of interest corresponding to estimated values [Para 0071, An agent may periodically verify the changes rate (e.g., percentage of the number of records changed from the entire dataset) in the properties information used for the training. If this rate hits a certain level, a new process for network training may be started and a set of neuronal networks may then be updated. The system may be scheduled to execute periodical analysis. This analysis can reflect future price increasing for certain types of properties or can reveal trends in the real estate market. When these results are available, the system can notify its clients or can push the results to the registered clients. The two sets of neuronal networks (e.g., any RVE neural networks (e.g., as described with respect to process 400) that may be used for obtaining an estimated property value (e.g., at step 505), which may be feed forward networks for estimation].
Packes teaches the above limitations of claim 14, but does not teach for each estimated value different from said first target value, obtaining a number of confusions corresponding to a number of times said estimated value has been estimated; and Increasing learning rates associated with values of interest correspond to estimated values with largest numbers of confusions.
Aceto teaches,
for estimated value different from target value, obtaining a number of confusions corresponding to a number of times estimated value has been estimated [Sect IV (B), para 4, To deepen the performance evaluation, Fig. 2 reports a calibration analysis, assessing whether the class-probability estimates are representative of the true-class (posterior) probabilities, where a miscalibrated classifier (on a given task) returns excessively optimistic or pessimistic decisions];
Aceto is analogous to the claimed invention as they both relate to deep learning. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Packes’ teachings to incorporate the teachings of Aceto and provide a confusion matrix [Aceto, Sect IV (B), para 4] in order to deepen the performance evaluation.
Packes-Aceto do not teach Increasing learning rates associated with values of interest correspond to estimated values with largest numbers of confusions.
Mittal teaches,
learning rates associated with values of interest correspond to estimated values with largest numbers of confusions [Para 0048, The disclosed implementation of BLT selects a confusing class in step 4 by using information from the confusion matrix C for a given tail class c… This strategy will select the most confusing classes more often… The disclosed implementation runs the gradient ascent image generation procedure with a learning rate δ=0.7. It stops running when S.sub.c′(I′)≥s.sub.c′ or when it reaches 15 iterations].
Mittal is analogous to the claimed invention as they both relate to identifying confusions. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Packes and Aceto’s teachings to incorporate the teachings of Mittal and provide learning rates associated with values of interest correspond to estimated values with largest numbers of confusions in order to provide a means for learning from misclassification in order to improve model results.
Packes-Aceto-Mittal do not teach increasing the learning rates.
Popescu teaches,
increasing the learning rates [Para 0172, During the first five epochs of training, the learning rate will increase linearly from 0 to 4×10.sup.−3].
Popescu is analogous to the claimed invention as they both relate to utilizing learning rates for training. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Packes, Aceto, and Mittal’s teachings to incorporate the teachings of Popescu and provide increasing learning rates in order to learn from desirable weights.
Regarding claim 15, Packes-Aceto teach the limitations of claim 14.
Packes further teaches,
Wherein: said properties are properties used to describe digital uses, said artificial neural network being configured to identify a digital use among a set of digital uses, each digital use being described by a digital behavior associated with at least one said property and by a digital environment associated with at least one said property [Para 0057, a user may utilize a website, a mobile app, an external application, and/or the like to specify any suitable attribute values for any suitable set of attributes. In one implementation, the user may specify attribute values for any of the attributes discussed with regard to step 105 of process 100. For example, the user may be enabled to enter attribute values for a minimum number of attributes (e.g., one or any other suitable number greater than one). In another example, the user may enter attribute values for a greater number of attributes to enhance the accuracy of the price prediction. The REP may be operative to enable a user to enter a minimum amount of information to allow the REP to identify a property (e.g., the address)];
said primary training comprising training said neural network to identify at least one target digital use among said set of digital uses, the data of said first set of primary training data being extracted from a first plurality of network packets captured during at least one execution of an application associated with said first target digital use, said primary training data being labeled by associating these data with a first set of target property values describing at least a digital behavior and a digital environment of said first target digital use; and the data of said set of secondary training data obtained for said at least one first target value being extracted from a second plurality of network packets captured during at least one execution of an application associated with a digital use whose said target property is described by said first target value [Para 0035, historical data may be prepared to reduce property value variation associated with time. For example, reducing property value variation associated with time (e.g., inflation, housing market trends, etc.) in historical data may lead to better results when training and/or retraining a neural network to estimate property value based on differences in attribute values. In one embodiment, to prepare historical data, historical property values (e.g., historical prices) may be analyzed to determine property values as of a specified (e.g., current) time period; Para 0040, Neural network parameters may be obtained at step 117 of process 100. In one implementation, the neural network parameters may be obtained from the administrator or any other suitable party via any suitable GUI of the REP. In another implementation, the neural network parameters may be obtained from a configuration file. All the parameters may be saved in a database. When a network is trained, all the information about this process may be saved in the database (e.g., the training parameters, the data set used, the network performances, training execution time, all the testing results for that specific network with results for each record, etc.); Para 0057, a user may utilize a website, a mobile app, an external application, and/or the like to specify any suitable attribute values for any suitable set of attributes; Para 0058, Attribute values for the property may be augmented at step 405 of process 400; Para 0094, the input/output devices may include one or more network devices 815… In one implementation, a network device may be a network card that may obtain (e.g., via a Category 5 Ethernet cable), process, output (e.g., via a wireless antenna), and/or the like network data (e.g., REP data)… A network card may be a… wireless modem based on cellular protocols such as high speed packet access (HSPA)].
Regarding claim 16, Packes further teaches,
A non-transitory computer-readable recording medium on which a computer program is recorded comprising instructions [Para 0026, a non-transitory computer-readable medium may include computer-readable instructions recorded thereon for training, with a processing system] which, when executed by a computer [Para 0027, a computer system-implemented method], cause the computer to implement the execution of
a method for training an artificial neural network so that said artificial neural network identifies at least one value of a property among a plurality of property values, each property being able to take at least two different values, said artificial neural network including an output layer including, for at least one said property value, a neuron configured to deliver a prediction score for said property value [Para 0020, the REP may train, retrain, and utilize sets of neural networks to estimate values of real estate properties and/or to predict values of real estate properties into the future… A neural network or neuronal network or artificial neural network may be hardware-based, software-based, or any combination thereof, such as any suitable model (e.g., a computational model), which, in some embodiments, may include one or more sets or matrices of weights; Para 0035, historical data may lead to better results when training and/or retraining a neural network to estimate property value based on differences in attribute values. In one embodiment, to prepare historical data, historical property values (e.g., historical prices) may be analyzed to determine property values as of a specified (e.g., current) time period. In one implementation, historical data may be prepared using the following preparation process. An estimation time period unit (e.g., one month) may be defined. Historical data may be obtained for a specified estimation time frame (e.g., the last year) for properties that have data regarding property values during the estimation time frame (e.g., properties that were sold during the last year and have a selling price, properties whose property values were evaluated during a previous preparation iteration, etc.). The obtained historical data may be sliced for each estimation time period (e.g., for each month during the last year); Para 0040, the neural network parameters may include the number of neurons in the output layer],
said method comprising: implementing a primary training comprising training said neural network to identify at least one target property value from a first set of primary training data labeled by associating these data with a first set of target property values [Para 0035, before it is used (e.g., as inputs, as outputs, etc.), historical data may be prepared to reduce property value variation associated with time. For example, reducing property value variation associated with time (e.g., inflation, housing market trends, etc.) in historical data may lead to better results when training and/or retraining a neural network to estimate property value based on differences in attribute values]; and
implementing a secondary training including: obtaining, for at least a first target value of a target property, a set of secondary training data; identifying an estimated value of the target property associated with at least one said secondary training data, by using the artificial neural network trained by said primary training [Para 0058, Attribute values for the property may be augmented at step 405 of process 400… if the user modifies an attribute value, the modified attribute value may replace the attribute value stored in the data store; Para 0020, the REP may train, retrain, and utilize sets of neural networks to estimate values of real estate properties and/or to predict values of real estate properties into the future… A neural network or neuronal network or artificial neural network may be hardware-based, software-based, or any combination thereof, such as any suitable model (e.g., a computational model), which, in some embodiments, may include one or more sets or matrices of weights];
learning rates (Para 0071, a new process for network training may be started) associated with neurons of the output layer of neural network that are associated with values of interest corresponding to estimated values [Para 0071, An agent may periodically verify the changes rate (e.g., percentage of the number of records changed from the entire dataset) in the properties information used for the training. If this rate hits a certain level, a new process for network training may be started and a set of neuronal networks may then be updated. The system may be scheduled to execute periodical analysis. This analysis can reflect future price increasing for certain types of properties or can reveal trends in the real estate market. When these results are available, the system can notify its clients or can push the results to the registered clients. The two sets of neuronal networks (e.g., any RVE neural networks (e.g., as described with respect to process 400) that may be used for obtaining an estimated property value (e.g., at step 505), which may be feed forward networks for estimation].
Packes teaches the above limitations of claim 1, but does not teach for each estimated value different from said first target value, obtaining a number of confusions corresponding to a number of times said estimated value has been estimated; and Increasing learning rates associated with values of interest correspond to estimated values with largest numbers of confusions.
Aceto teaches,
for estimated value different from target value, obtaining a number of confusions corresponding to a number of times estimated value has been estimated [Sect IV (B), para 4, To deepen the performance evaluation, Fig. 2 reports a calibration analysis, assessing whether the class-probability estimates are representative of the true-class (posterior) probabilities, where a miscalibrated classifier (on a given task) returns excessively optimistic or pessimistic decisions.];
Aceto is analogous to the claimed invention as they both relate to deep learning. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Packes’ teachings to incorporate the teachings of Aceto and provide a confusion matrix [Aceto, Sect IV (B), para 4] in order to deepen the performance evaluation.
Packes-Aceto do not teach Increasing learning rates associated with values of interest correspond to estimated values with largest numbers of confusions.
Mittal teaches,
learning rates associated with values of interest correspond to estimated values with largest numbers of confusions [Para 0048, The disclosed implementation of BLT selects a confusing class in step 4 by using information from the confusion matrix C for a given tail class c… This strategy will select the most confusing classes more often… The disclosed implementation runs the gradient ascent image generation procedure with a learning rate δ=0.7. It stops running when S.sub.c′(I′)≥s.sub.c′ or when it reaches 15 iterations].
Mittal is analogous to the claimed invention as they both relate to identifying confusions. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Packes and Aceto’s teachings to incorporate the teachings of Mittal and provide learning rates associated with values of interest correspond to estimated values with largest numbers of confusions in order to provide a means for learning from misclassification in order to improve model results.
Packes-Aceto-Mittal do not teach increasing the learning rates.
Popescu teaches,
increasing the learning rates [Para 0172, During the first five epochs of training, the learning rate will increase linearly from 0 to 4×10.sup.−3].
Popescu is analogous to the claimed invention as they both relate to utilizing learning rates for training. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Packes, Aceto, and Mittal’s teachings to incorporate the teachings of Popescu and provide increasing learning rates in order to learn from desirable weights.
Claim(s) 3 is/are rejected under 35 U.S.C. 103 as being unpatentable over Packes in view of Aceto, Mittal, and Popescu, and in further Manoharan et al. (US 20210303998 A1), hereinafter Manoharan.
Packes-Aceto-Mittal-Popescu do not teach during said primary training, said learning rates associated with said neurons of the output layer of said neural network have the same value.
Manoharan teaches,
during said primary training, said learning rates associated with said neurons of the output layer of said neural network have the same value [Here, the layers in the neural network are divided into groups—initial layers, layers in the middle and final layers. The same learning rate has been used for each layer in a group… the layers in the middle and the final layers use the same learning rates].
Manoharan is analogous to the claimed invention as they both relate to reinforcement learning methodologies. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Packes, Aceto, Mittal, and Popescu‘s teachings to incorporate the teachings of Manoharan and provide same learning rates in order to iteratively achieve similar outcomes.
Claim(s) 4 is/are rejected under 35 U.S.C. 103 as being unpatentable over Packes in view of Aceto, Mittal, and Popescu, and in further view of Rouzbeh et al. (EP 2386987 A1, see attached translation), hereinafter Rouzbeh.
Regarding claim 4, Packes-Aceto-Mittal-Popescu teach the limitations of claim 1 including said increase of said learning rate (Popescu, Para 0172).
Packes-Aceto-Mittal-Popescu do not teach multiplying learning rate by a constant.
Rouzbeh teaches,
multiplying learning rate by a constant [Para 0028, the agent increases the learning rate through multiplication by a predetermined factor].
Rouzbeh is analogous to the claimed invention as they both relate to reinforcement learning methodologies. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Packes, Aceto, Mittal, and Popescu‘s teachings to incorporate the teachings of Rouzbeh and provide multiplying learning rate by a constant in order to create a diverse training set that improves expected results.
Claim(s) 11 is/are rejected under 35 U.S.C. 103 as being unpatentable over Packes in view of Aceto, Mittal, and Popescu, and in further view of Lin et al. (US 20210367885 A1), hereinafter Lin.
Regarding claim 11, Packes-Aceto-Mittal-Popescu teach the limitations of claim 2 including obtaining said first set of primary training data and obtaining said first set of secondary training data (claim 1: Packes, paras 0035 and 0057).
Packes-Aceto-Mittal-Popescu do not teach obtaining a subset of network packets and, for each network packet of said subset: obtaining a set of data of the network packet, and processing said set of data of the network packet, so as to obtain the first set of training data, the processing comprising: for each data of said set of data of the network packet taking the form of a categorical variable, converting the value into a vector of binary values, and for each data of said set of data of the network packet having a numerical value set in an interval different from the interval comprised between 0 and 1, normalizing the value so that the value is set in the interval comprised between 0 and 1.
Lin teaches
obtaining a subset of network packets [Para 0006, retrieving a plurality of testing data transmitted between the monitored network and the external network through the network interface] and, for each network packet of said subset:
obtaining a set of data of the network packet [Para 0006, retrieving a plurality of testing data transmitted between the monitored network and the external network through the network interface; preprocessing a plurality of packet headers of the pluralities of testing data], and
processing said set of data of the network packet, so as to obtain set of training data [Para 0007, choosing the packet headers of the pluralities of training data with certain feature types, in which the feature types comprises numerical features and non-numerical features; filling up fields of missing feature values of the feature types; standardizing the numerical features in the feature types to obtain first sub-features;], the processing comprising:
for each data of said set of data of the network packet taking the form of a categorical variable, converting the value into a vector of binary values [Para 0007, converting the non-numerical features in the feature types to binary value features to obtain second sub-features], and
for each data of said set of data of the network packet having a numerical value set in an interval different from the interval comprised between 0 and 1, normalizing the value so that the value is set in the interval comprised between 0 and 1 [Para 0007, converting the non-numerical features in the feature types to binary value features to obtain second sub-features; and obtaining the training feature vectors according to the first sub-features and the second sub-features; The step of S303 may further comprise the steps of standardizing the numerical features in the packet headers of the pluralities of training data to obtain first sub-features by the processor 101 (S405), for example, using the min-max normalization to process the features in training data.].
Lin is analogous to the claimed invention as they both relate to processing packets to create training data. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Packes, Aceto, Mittal, and Popescu‘s teachings to incorporate the teachings of Lin and provide normalizing the training data in order to improve model accuracy, speed up training convergence, and reduce overfitting.
Claim(s) 12 is/are rejected under 35 U.S.C. 103 as being unpatentable over Packes in view of Aceto, Mittal, Popescu, and Lin, and in further view of Luo et al. (CN 113516330 A, see attached translation), hereinafter Luo.
Regarding claim 12, Packes-Aceto-Mittal-Popescu-Lin teach the limitations of claim 12.
Packes-Aceto-Mittal-Popescu-Lin do not teach wherein the subset of network packets comprises a maximum of one hundred network packets.
Luo teaches,
wherein subset of network packets comprises a maximum of one hundred network packets [Para 0002, users can send virtual resources to each other. Typically, virtual resources are stored in multiple resource packages, and resource links corresponding to the multiple resource packages are created and displayed to users. Users trigger the resource links to request resource packages; Para 0143, if the first reference quantity range is [0.1, 100], it means that the minimum value of the unit resource quantity corresponding to each sub-resource package is 0.1, and the maximum value is 100].
Luo is analogous to the claimed invention as they both relate to network packet acquisition and processing. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Packes, Aceto, and Lin’s teachings to incorporate the teachings of Luo and provide a constraint on the maximum allowed network packages to speed up model convergence.
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.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SYED RAYHAN AHMED whose telephone number is (571)270-0286. The examiner can normally be reached Mon-Fri ET.
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, David Yi can be reached at (571) 270-7519. 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.
/SYED RAYHAN AHMED/Examiner, Art Unit 2126
/DAVID YI/Supervisory Patent Examiner, Art Unit 2126
Prosecution Insights