DETAILED ACTION
1. This office action is in response to the Application No. 18740960 filed on 02/04/2026. Claims 1-20 are presented for examination and are currently pending.
Notice of Pre-AIA or AIA Status
2. The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Continued Examination Under 37 CFR 1.114
3. A request for continued examination under 37 CFR 1.114, including the fee set
forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this
application is eligible for continued examination under 37 CFR 1.114, and the fee set
forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action
has been withdrawn pursuant to 37 CFR 1.114. Applicant’s submission filed on
02/04/2026 has been entered.
Response to Arguments
4. The arguments on page 14 is persuasive because as argued by the Applicant that the amended claims recite a specific training and deployment architecture that improves how neural networks are constructed, trained, and deployed by: (i) selecting an intermediate network representation based on information value rankings of predictive power; (ii) extracting a time-indexed hidden-state vector from an upper hidden layer as a supervisory target rather than relying on final output layers; (iii) enforcing bounded-range compatibility between a teacher network's internal representations and a student network's outputs; and (iv) decoupling training from deployment to enable an unbounded-complexity teacher model while deploying a smaller, interpretable, and computationally efficient student model in an existing system architecture. Also, this architecture improves computer functionality by enabling more efficient, stable, and deployable predictive models- reducing computational and deployment burdens while preserving predictive performance. The claimed invention, a specific computer-implemented mechanism for selecting, extracting, and transferring internal neural-network representations to improve model deployment and operation. This leads to an improvement in computer technology. As a result, the 101 rejection has been withdrawn.
The Applicant’s arguments have been considered but are moot in view of the
new grounds of rejection. The Examiner is withdrawing the rejections in the previous
office action 12/05/2025 because the applicant amendments necessitated the new
grounds of rejection presented in this office action. Lan in view of Ward has been applied to teach the independent claims.
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.
5. Claims 1-3, 5, 8-10, 12, 15, 16 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Lan et al. (US20190080176 PCT filed 03/28/2017) in view of Ward et al. (US10210860 filed 08/22/2018)
Regarding claim 1, Lan teaches a method comprising: obtaining, with at least one processor, first training data associated with a first set of features (The dataset 112 may include one or more labeled videos. One or more of the frames in the videos may be labeled as being associated with a predefined action [0018]) and second training data associated with a second set of features (input frame 122 [0022]; input frame being associated with a predefined action label indicating the frame includes no action [0074]) different than the first set of features (The dataset 112 may include one or more labeled videos. One or more of the frames in the videos may be labeled as being associated with a predefined action [0018]. The Examiner notes the dataset 112 include predefined action while input frame 122 includes no action);
training, with the at least one processor, a first neural network (learning sub-network 310 [0074], Fig. 8) based on the first training data (training dataset 112 [0078]; the parameters for the feature learning sub-network 310 … may be determined based on the training dataset [0044]) and the second training data (input frame 122 [0022]; In some implementations, when receiving representation information of an incoming frame of a video, the feature learning sub-network 310 may extract features for the frame based on the representation information [0052]; … frames among the predefined action labels based on the features from the last layer of the feature learning sub-network 310 [0042]),
the first neural network (learning sub-network 310 [0074], Fig. 8) including an input layer (the RNN layer 312 [0046], Fig. 8), an output layer (A further SoftMax layer 340 is added to receive the outputs of features of an input frame from the subsequent FC layer 318 [0074], Fig. 8. The Examiner notes SoftMax layer 340 is the output layer in Fig. 8), and
a plurality of hidden layers (The feature learning sub-network 310 may also include one or more FC layers, for example, FC layers 314 and 318, each including one or more neurons [0040]) including an intermediate layer (FC layers 318 [0040], Fig. 8) between the input layer (RNN layer 312 [0046], Fig. 8) and the output layer (SoftMax layer 340, Fig. 8); and
training, with at least one processor, a second neural network (Additionally, the learning network 500 includes a regression sub-network 330 to perform the forecast task. The feature learning sub-network 310 may be shared by the … regression sub-network 330 [0057], Fig. 8. The Examiner notes sub-network 330 as the second neural network) based on the second training data (the regression sub-network 330 may include a FC layer 336 to determine confidence(s) of the input frame being the frame at the start point of an action and/or the frame at the end point of the action [0064]. The Examiner notes input frame is the second training data)
and using a loss function (the objective function for … the separate training of the sub-networks 310 is to reduce or minimize the maximum likelihood loss function [0048]) that depends directly on a hidden state vector extracted directly from the intermediate layer (layer 318 [0074], Fig. 8; The Examiner notes layer 318 is a hidden layer and softmax layer 340 is an output layer because the Applicant discloses in instant specification that “output or final layer of a neural network is a softmax layer” and “output of the intermediate layer of the first model includes a hidden state vector” (instant specification: [0113]) of the first neural network (learning sub-network 310 [0074], Fig. 8) without passing through the output layer (SoftMax layer 340, Fig. 8) of the first neural network (learning sub-network 310 [0074], Fig. 8) and an output (FC layer 336, Fig. 8) of the second neural network (sub-network 330, Fig. 8),
Lan does not explicitly teach wherein the first set of features includes complex features and the second set of features includes interpretable features, wherein the first neural network includes a greater number of parameters than the second neural network, and wherein training the first neural network is decoupled from deployment of the second neural network such that a complexity of the first neural network is unbounded relative to the second neural network; determining, with the at least one processor, a plurality of information values of a plurality of intermediate layers of the first neural network, and selecting the intermediate layer from the plurality of intermediate layers based on the plurality of information values, the plurality of information values ranking predictive power of outputs of the plurality of intermediate layers, wherein the first neural network includes at least two hidden layers configured to generate hidden state vectors as functions of time, the selected intermediate layer comprises an upper hidden layer of the at least two hidden layers, and the hidden state vector comprises a hidden state vector h(t) extracted from the upper hidden layer and serving as a target vector for training the second neural network, wherein the hidden state vector h(t) is bounded within a finite output range because the hidden state vector h(t) is calculated as h(t) = o(t) O tanh(c(t)), where o(t) is an output from an output gate and c(t) is a cell status, and wherein the second neural network includes an activation function configured to constrain the output of the second neural network to be bounded within a finite output range corresponding to the finite output range of the hidden state vector h(t);providing, with the at least one processor, the trained second neural network for deployment in an existing system architecture; obtaining, with the at least one processor, input data; and processing, with the at least one processor and the trained second neural network, the input data to generate an output comprising a prediction for the input data.
Ward teaches wherein the first set of features includes complex features (features of the input audio stream, col. 5, ln 4-5) and the second set of features includes interpretable features (groups of features as words, roughly similar to an acoustic model and a pronunciation dictionary, col. 5, ln 5-7), wherein the first neural network includes a greater number of parameters (As the complexity of the decision boundary exceeds the size of what can be easily expressed in the size of the model, various components of the neural network, such as weights and hidden nodes, become overloaded, col. 19, ln 53-56) than the second neural network (First fully-connected layer 203 may resize the output of CNN stack 202 for consumption by the subsequent stack ... In some embodiments, the first fully-connected layer 203 may reduce the dimension of this output to reduce the number of parameters subsequent stack need to process, col. 9, ln 37-44), and wherein training the first neural network is decoupled from deployment of the second neural network such that a complexity of the first neural network is unbounded relative to the second neural network (A fully-connected neural network is one kind of densely connected neural network, where a densely connected neural network is one where most of the nodes in each layer of the neural network have edge connections to most of the nodes in the subsequent layer. The aforementioned definitions may exclude the output layer which has no outbound connections, col. 8, ln 63-67 to col. 9, ln 1-3);
determining, with the at least one processor (The apparatuses and methods described in this application may be partially or fully implemented by one or more computer programs executed by one or more processors, col. 4, ln 23-25),
a plurality of information values of a plurality of intermediate layers of the first neural network (While recurrent neural network 502 is illustrated as a single layer for the purposes of illustration, it is to be understood that the recurrent network may include any number of layers. For each time step, recurrent neural network 502 produces a set of features related to a word prediction 503a-n at that time step. This set of features is expressed as a tensor or vector output and is directly input to subsequent layers. The Examiner notes the features expressed as a tensor are information values), and
selecting the intermediate layer from the plurality of intermediate layers based on the plurality of information values (In some embodiments, a whole layer of the neural network may be selected for the internal state representation, pg. 35, ln 26-27),
the plurality of information values ranking predictive power of outputs of the plurality of intermediate layers (In an embodiment, at each iteration, the bandit algorithm selects a general training subset to train on by applying the scoring function to each subset and choosing the highest scoring one, pg. 28, ln 45-48),
wherein the first neural network includes at least two hidden layers configured to generate hidden state vectors as functions of time (LSTM and GRU type RNNs include at least one back loop where the output activation of a neural network enters as an input to the neural network at the next time step. In other words, the output activation of at least one neural network node is an input to at least one neural network node of the same or a prior layer in a successive time step, col. 10, ln 28-33),
the selected intermediate layer comprises an upper hidden layer of the at least two hidden layers, and the hidden state vector comprises a hidden state vector h(t) extracted from the upper hidden layer (Example neural network 1100 is a fully-connected neural network with multiple layers of hidden states. Neural network layer portion 1110 is a selector (col. 20, ln 7-10); Although Neural Network Memory has been illustrated in fully-connected neural networks 1100, 1150 it may be used in any other form of neural network, such as ... RNN layers 204, 704, 804, col. 21, ln 48-52)
and serving as a target vector for training the second neural network (The target value for training examples is the neural network node weights in the general model, col. 23, ln 56-58),
wherein the hidden state vector h(t) is bounded within a finite output range because the hidden state vector h(t) is calculated as h(t) = o(t) O tanh(c(t)), where o(t) is an output from an output gate and c(t) is a cell status (The LSTM may comprise an output gate layer comprising a neural network layer with a sigmoid activation function input to a pointwise multiplication gate with the other input being the hidden state after being passed through the tan h function. The result of this operation may be output as the tensor output of the LSTM at the current time step, col. 10, ln 59-65), and
wherein the second neural network includes an activation function configured to constrain the output of the second neural network to be bounded within a finite output range corresponding to the finite output range of the hidden state vector h(t) (One method of deriving a low-precision feature is to quantize the output of an activation function of a node of a neural network. For example, in an embodiment, the output of the activation function at each node of the portion may be simplified into a binary representation. That is, any output of the node above a threshold is treated as a first binary value, and any output of the node below the threshold is treated as a second binary value, col. 34, ln 58-65);
providing, with the at least one processor (The apparatuses and methods described in this application may be partially or fully implemented by one or more computer programs executed by one or more processors, col. 4, ln 23-25),
the trained second neural network for deployment in an existing system architecture (a neural network is trained on general training set 1510 to create a general model and different mixes of general training subsets 1511-1515 are used for further training to customize the neural network, col. 26, ln 29-32);
obtaining, with the at least one processor, input data (The training subsets 1511-1515 may have different characteristics, such as source (e.g., public dataset, ...), ... audio quality (e.g., phone conversations, in-person recordings, speaker phones), and so on (col. 26, ln 6-13); The set of general training subsets 1511-1515 used for training may be adjusted to improve performance on the custom evaluation subset 1522, col. 26, ln 26-28); and
processing, with the at least one processor (The apparatuses and methods described in this application may be partially or fully implemented by one or more computer programs executed by one or more processors, col. 4, ln 23-25)
and the trained second neural network, the input data to generate an output comprising a prediction for the input data (At each iteration, it a selects general training subset 1511-1515 to train on. The neural network is trained on the selected general training subset for a number of training batches, where the number of training batches may be configurable (col. 27, ln 37-41); For each spoken word in the input audio, one frame of the output sequence will be desired to have a high probability prediction for a word of the vocabulary, col. 12, ln 4-7).
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to have modified the method of Lan to incorporate the teachings of Ward for the benefit of a process that continues indefinitely to iteratively improve the neural network's performance (col. 27, ln 53-54) and increasing accuracy (Ward, col. 3, ln 5)
Regarding claim 2, Lan and Ward teaches the method of claim 1, Lan teaches wherein the first neural network includes at least one of the following: a deep neural network, a recurrent neural network, an ensemble of a plurality of neural networks, or any combination thereof (The feature learning sub-network 310 may include one or more RNN layers, for example, RNN layers 312 and 316, each including one or more LSTM neurons (such as LSTM neurons 200) [0040]).
Regarding claim 3, Lan and Ward teaches the method of claim 2, Lan teaches wherein the second neural network includes a feedforward regression neural network (FRNN) (In some implementations, the regression sub-network 330 may be trained to determine the parameters for the neuron(s) in the FC layers so that the output(s) of the FC layers can indicate the confidence(s) [0064]).
Regarding claim 5, Lan and Ward teaches the method of claim 2, Lan teaches wherein the plurality of hidden layers includes a plurality of long short-term memory (LSTM) hidden layers including the intermediate layer of the first neural network (the feature learning sub-network 310 may include only one RNN layer including a plurality of LSTM neurons [0041])
Regarding claim 8, claim 8 is similar to claim 1. It is rejected in same manner and reasoning applying.
Regarding claim 9, claim 9 is similar to claim 2. It is rejected in same manner and reasoning applying.
Regarding claim 10, claim 10 is similar to claim 3. It is rejected in same manner and reasoning applying.
Regarding claim 12, claim 12 is similar to claim 5. It is rejected in the same manner and reasoning applying.
Regarding claim 15, claim 15 is similar to claim 1. It is rejected in same manner and reasoning applying. Further, Lan teaches a computer program product comprising at least one non-transitory computer-readable medium including program instructions that, when executed by at least one processor, cause the at least one processor to (These program codes may be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the program codes, when executed by the processor or controller, cause the functions/operations specified in the flowcharts and/or block diagrams to be implemented [0098]):
Regarding claim 16, claim 16 is similar to claim 2. It is rejected in same manner and reasoning applying.
Regarding claim 19, claim 19 is similar to claim 5. It is rejected in same manner and reasoning applying.
6. Claims 4, 11 are rejected under 35 U.S.C. 103 as being unpatentable over Lan et al. (US20190080176 PCT filed 03/28/2017) in view of Ward et al. (US10210860 filed 08/22/2018) in view of Huang et al (US20180365564) and further in view of Caelen et al. (US20200257964 filed 07/13/2018)
Regarding claim 4, Lan and Ward teaches the method of claim 3, Lan teaches wherein the second neural network (Additionally, the learning network 500 includes a regression sub-network 330 to perform the forecast task. The feature learning sub-network 310 may be shared by the … regression sub-network 330 [0057], Fig. 8. The Examiner notes sub-network 330 as the second neural network) includes at least one first layer (the regression sub-network 330 may include a FC layer 332 for feature fusion [0062]) and at least one second layer, wherein the output of the second neural network includes an output of the at least one first layer (The features output from the FC layer 332 may be processed in the soft selector 334 [0071]. The Examiner notes that soft selector 334 as the second layer),
Lan and Ward does not explicitly teach wherein the at least one first layer of the second neural network includes a regression neural network, and wherein the at least one second layer of the second neural network includes a logistic regression model.
Huang wherein the at least one first layer of the second neural network includes a regression neural network (when the task of the student network is a regression task, the form of the task specific loss function is a distance loss function [0072]; the features of the second middle layer refer to feature maps output from a second specific network layer of the student network after the training sample data are provided to the student network [0068]; the second specific network layer is a middle network layer or the last network layer of the student network [0069]), and
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to have modified the method of Lan and Ward to incorporate the teachings of Huang for the benefit of transferring knowledge of features of a middle layer of the teacher network to the student network (Huang, abstract)
Modified Lan does not explicitly teach wherein the at least one second layer of the second neural network includes a logistic regression model.
Caelen teaches wherein the at least one second layer of the second neural network includes a logistic regression model (The distribution over classes fraud and non-fraud given state st is modeled with a logistic regression output model [0041])
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to have modified the method of Modified Lan to incorporate the teachings of Caelen for the benefit of taking into account the time elapsed between two authentication, operation or transactions (Caelen, [0017])
Regarding claim 11, claim 11 is similar to claim 4. It is rejected in the same manner and reasoning applying.
7. Claims 6, 13 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Lan et al. (US20190080176 PCT filed 03/28/2017) in view of Ward et al. (US10210860 filed 08/22/2018) and further in view of Mequanint et al. US 20200218878 filed 1/3/2019
Regarding claim 6, Lan and Ward teaches the method of claim 1, Lan teaches wherein the loss function used to train the second neural network (the objective function for training the regression sub-network 330 is to reduce or minimize the loss function for each input frame [0065]) is different than a loss function used to train the first neural network (the objective function for … the separate training of the sub-networks 310 is to reduce or minimize the maximum likelihood loss function [0048]), and
… before the second neural network is trained using the loss function (the objective function for training the regression sub-network 330 is to reduce or minimize the loss function for each input frame [0065]) that depends directly on the hidden state vector extracted directly from the intermediate layer (layer 318 [0074], Fig. 8; The Examiner notes layer 318 is a hidden layer and softmax layer 340 is an output layer because the Applicant discloses in instant specification that “output or final layer of a neural network is a softmax layer” and “output of the intermediate layer of the first model includes a hidden state vector” (instant specification: [0113]) of the neural network (learning sub-network 310 [0074], Fig. 8) without passing through the output layer (SoftMax layer 340, Fig. 8) of the first neural network (learning sub-network 310 [0074], Fig. 8) and the output (FC layer 336, Fig. 8) of the second neural network (sub-network 330, Fig. 8).
Lan and Ward does not explicitly teach wherein the first neural network is trained separately offline
Mequanint teaches wherein the first neural network is trained separately offline (In some cases, the neural network may be trained off device (e.g., offline on some different server, on a local machine/computer, etc.) [0026])
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to have modified the method of Modified Lan to incorporate the teachings of Mequanint for the benefit of reducing computationally intensive training of the neural network, reducing large amounts of memory, processing, power, and time (Mequanint [0037])
Regarding claim 13, claim 13 is similar to claim 6. It is rejected in same manner and reasoning applying.
Regarding claim 20, claim 20 is similar to claim 6. It is rejected in same manner and reasoning applying.
8. Claims 7, 14 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Lan et al. (US20190080176 PCT filed 03/28/2017) in view of Ward et al. (US10210860 filed 08/22/2018) in view of Rybakov et al. (US10824940 filed 11/30/2016)
Regarding claim 7, Lan and Ward teaches the method of claim 1, Lan teaches wherein the loss function minimizes the loss (the objective function for … the separate training of the sub-networks 310 is to reduce or minimize the maximum likelihood loss function [0048]) between the output of hidden state vector extracted directly from the intermediate layer (layer 318 [0074], Fig. 8; The Examiner notes layer 318 is a hidden layer and softmax layer 340 is an output layer because the Applicant discloses in instant specification that “output or final layer of a neural network is a softmax layer” and “output of the intermediate layer of the first model includes a hidden state vector” (instant specification: [0113]) of the first neural network (learning sub-network 310 [0074], Fig. 8) without passing through the output layer (SoftMax layer 340, Fig. 8) of the first neural network (learning sub-network 310 [0074], Fig. 8) and the output (FC layer 336, Fig. 8) of the second neural network (sub-network 330, Fig. 8), and
wherein the loss function does not depend on the first training data associated with the first set of features (the objective function for training the regression sub-network 330 is to reduce or minimize the loss function for each input frame [0065]).
Lan and Ward does not explicitly teach wherein the loss function minimizes the squared error (L2) loss
Rybakov teaches wherein the loss function minimizes the squared error (L2) loss (The artificial neural network 200A, 200B, 200C may also use a cost function to find an optimal solution (e.g., an optimal activation function). ... In an example, the cost function includes a mean-squared error function that minimizes the average squared error between an output ƒ (x) and a target value y over the example pairs (x, y), col. 8, lines 60-67)
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to have modified the method of Lan and Ward to incorporate the teachings of Rybakov for the benefit of the neural network temporal ensemble which can produce a recommendation with accuracy (Rybakov, col 2, lines 58-60)
Regarding claim 14, claim 14 is similar to claim 7. It is rejected in same manner and reasoning applying.
Regarding claim 17, Lan and Ward teaches the computer program product of claim 16, Lan teaches wherein the second neural network includes a feedforward regression neural network (FRNN) (In some implementations, the regression sub-network 330 may be trained to determine the parameters for the neuron(s) in the FC layers so that the output(s) of the FC layers can indicate the confidence(s) [0064]), and
wherein the loss function minimizes the … loss (the objective function for … the separate training of the sub-networks 310 is to reduce or minimize the maximum likelihood loss function [0048]) between the hidden state vector extracted directly from the intermediate layer (layer 318 [0074], Fig. 8; The Examiner notes layer 318 is a hidden layer and softmax layer 340 is an output layer because the Applicant discloses in instant specification that “output or final layer of a neural network is a softmax layer” and “output of the intermediate layer of the first model includes a hidden state vector” (instant specification: [0113]) of the first neural network (learning sub-network 310 [0074], Fig. 8) without passing through the output layer (SoftMax layer 340, Fig. 8) of the first neural network (learning sub-network 310 [0074], Fig. 8) and the output (FC layer 336, Fig. 8) of the second neural network (sub-network 330, Fig. 8), and
wherein the loss function does not depend on the first training data associated with the first set of features (the objective function for training the regression sub-network 330 is to reduce or minimize the loss function for each input frame [0065]).
Lan and Ward does not explicitly teach wherein the loss function minimizes the squared error (L2) loss
Rybakov teaches wherein the loss function minimizes the squared error (L2) loss (The artificial neural network 200A, 200B, 200C may also use a cost function to find an optimal solution (e.g., an optimal activation function). ... In an example, the cost function includes a mean-squared error function that minimizes the average squared error between an output ƒ (x) and a target value y over the example pairs (x, y), col. 8, lines 60-67)
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to have modified the method of Lan and Ward to incorporate the teachings of Rybakov for the benefit of the neural network temporal ensemble which can produce a recommendation with accuracy (Rybakov, col 2, lines 58-60)
9. Claims 18 are rejected under 35 U.S.C. 103 as being unpatentable over Lan et al. (US20190080176 PCT filed 03/28/2017) in view of Ward et al. (US10210860 filed 08/22/2018) in view of Rybakov et al. (US10824940 filed 11/30/2016) in view of Huang et al (US20180365564) and further in view of Caelen et al. (US20200257964 filed 07/13/2018)
Regarding claim 18, Modified Lan teaches the computer program product of claim 17, Lan teaches wherein the second neural network (Additionally, the learning network 500 includes a regression sub-network 330 to perform the forecast task. The feature learning sub-network 310 may be shared by the … regression sub-network 330 [0057], Fig. 8. The Examiner notes sub-network 330 as the second neural network) includes at least one first layer (the regression sub-network 330 may include a FC layer 332 for feature fusion [0062]) and at least one second layer, wherein the output of the second neural network includes an output of the at least one first layer, (The features output from the FC layer 332 may be processed in the soft selector 334 [0071]. The Examiner notes that soft selector 334 as the second layer)
Modified Lan does not explicitly teach wherein the at least one first layer of the second neural network includes a regression neural network, and wherein the at least one second layer of the second neural network includes a logistic regression model.
Huang teaches wherein the at least one first layer of the second neural network includes a regression neural network (when the task of the student network is a regression task, the form of the task specific loss function is a distance loss function [0072]; the features of the second middle layer refer to feature maps output from a second specific network layer of the student network after the training sample data are provided to the student network [0068]; the second specific network layer is a middle network layer or the last network layer of the student network [0069]), and
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to have modified the method of Modified Lan to incorporate the teachings of Huang for the benefit of transferring knowledge of features of a middle layer of the teacher network to the student network (Huang, abstract)
Modified Lan does not explicitly teach wherein the at least one second layer of the second neural network includes a logistic regression model.
Caelen teaches wherein the second layer of the second model includes a logistic regression model (The distribution over classes fraud and non-fraud given state st is modeled with a logistic regression output model [0041])
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to have modified the method of Modified Lan to incorporate the teachings of Caelen for the benefit of taking into account the time elapsed between two authentication, operation or transactions (Caelen, [0017])
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MORIAM MOSUNMOLA GODO whose telephone number is (571)272-8670. The examiner can normally be reached Monday-Friday 8:00am-5:00pm EST.
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, Michelle T. Bechtold can be reached on (571) 431-0762. 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.
/M.G./Examiner, Art Unit 2148
/MICHELLE T BECHTOLD/Supervisory Patent Examiner, Art Unit 2148
Prosecution Insights