CTFR 16/818,266 CTFR 85199 DETAILED ACTION 07-03-aia AIA 15-10-aia The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA. Response to Arguments Applicant's arguments with respect to the rejections of claims 21, 28, and 35 and their respective dependent claims have been considered but they are moot in view of new grounds of rejections. As to any arguments not specifically addressed, they have been discussed in previous Office actions or are disclosed in the Claim Rejections section 103 below. Response to Amendment Claim Rejections - 35 USC § 103 07-103 AIA 3. The text of those sections of Title 35, U.S. Code not included in this section can be found in a prior Office action. 4. Claims 21, 28, and 35 are rejected under AIA 35 U.S.C. 103 as being unpatentable over Jaderberg et al. (US Publication 2021/0004676) in view of Prabhakara et al. (US Patent 12,468,960). Regarding claim 21, Jaderberg discloses one or more processors comprising: circuits to cause one or more learning rates of one or more neural networks to be adjusted based, at least in part, on input, wherein the input is based at least in part, on comparisons between a first output and a second output of the one or more neural networks, wherein the first output is produced by the one or more neural networks using a first set of weights and the second output is produced by the one or more neural networks using a second set of weights ( Jaderberg, fig. 3, para’s 0097-0136, the system executes iterative training processes for network A 320A and network B 320B, in parallel (indicated by line connectors 350A and 350B, respectively). As the system 100 executes the iterative training processes for network A 320A and network B 320B, network parameters A 325A and network parameters B 325B are updated and replaced by updated network parameters A 340A and updated network parameters B 340B, respectively. The system 100 determines an updated quality measure B 336B for the network B 320B with the values of the hyperparameters B 330B and the updated network parameters B 340B, (i.e., the system 100 executes step 204 of process 200 for network B 320B, represented as line connector 360 in FIG. 3). The system 100 determines new hyperparameters B 331B and new network parameters B 341B for the network B 320B (i.e., the system 100 executes step 206 of process 200 for network B 320B); para’s 0100-0101, the system 100 compares the quality measure A 335A and the quality measure B 335B corresponding to the respective output of network A 320A and B 320B as described in para. 0060, and sets new network parameters B 341B to be equal to the updated network parameters A 340A (represented as line connector 370), because in this example, network A 320A is selected as having better than network B 320B as a result of the comparison. The system 100 determines new hyperparameters B 341B, i.e., learning rate as described in para. 0053, for the network B 320B, represented by line connector 380 ; the system 100 determines from the new network parameters B 341B and the new hyperparameters B 331B for the network B 320B, a new quality measure B 337B (i.e., the system 100 executes step 208 of process 200 for the network B 320B). Jaderberg does not explicitly disclose but Prabhakara discloses the input indicating a selection by one or more users, wherein the user input is based at least in part, on one or more user comparisons between a first output and a second output of the one or more neural networks, presented to the one or more users using a user interfac e ( Prabhakara, col. 3 lines 43-56, model builder 212 determines a best set of hyperparameters by determining/selecting a best model trained and tested using the training data set and the test data set of the historical data. The best set of hyperparameters is used to train a prediction model with the full historical data set. The output of the prediction model is compared to a forecast that is input via forecast interface 204 of interface 202. The comparison is used to identify anomalies and these are provided to the user via feedback module 214 of applications 210 and feedback interface 206 of interface 202. User feedback indicating valid detected anomalies and anomalies not detected is used to retrain the prediction model using model builder 21 2). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to incorporate Prabhakara’s features into Jaderberg’s invention for improving the accuracy of the learning model by using user’s evaluations of the output of the models as data for tuning the model. 5. Claims 21, 22, 24, 26, 28-29, 31, 33, 35-37 and 39 are rejected under AIA 35 U.S.C. 103 as being unpatentable over Teig et al. (US Patent 11,610,154) in view of Prabhakara et al. (US Patent 12,468,960). Regarding claim 21, Teig discloses one or more processors comprising: circuits to cause one or more learning rates of one or more neural networks to be adjusted based, at least in part, on input, wherein the input is based at least in part, on comparisons between a first output and a second output of the one or more neural networks, wherein the first output is produced by the one or more neural networks using a first set of weights and the second output is produced by the one or more neural networks using a second set of weights ( Teig, col. 9 line 40 through col. 11 line 63, for each training run, the training system adjusts the weight value to generate predicted output; the error calculator 310 of the training system computes the error for the run of each input as the network 325 generates its output. The error calculator 310 receives both the predicted “ground truth” output from the input generator 305 and the output of the network 325, and uses a loss function that quantifies the difference between the predicted output and the actual output for each input; therefore, for each respective run, the loss function is quantified by measuring the respective predicted output with the actual output; the quantification can also be interpreted as comparing the output distance from one run with the output distance from another with respect to an actual output, i.e. a baseline comparison ; the training system performs multiple training runs with changing training inputs 335 , and performs validation using the validation system 350 to determine how predictive the network parameters are after each training run ; each training run In addition, the validation system 350 is used to modify the training parameters 340 in order to optimize the resulting network. As shown, the validation system 350 includes an input generator 355, a network 360, an error calculator 365, a description length score 370, and a hyperparameter modifier 375.; the validation system receives the weight values 330 (and any other parameters of the network 360) as trained by the training system 300 and measures the predictiveness of this network . The network 360 has the same structure as the network 325 used for training, and is used to validate the training by determining how predictive the weight values 330 are for validation inputs 380 ; the error calculator 365 calculates the error in the network output for the validation inputs 380 , in order to measure the predictiveness of the network after a training run . A description length score calculator 370 uses the measured error in some embodiments, along with additional information (e.g., possible hyperparameter modifications, calculations of error due to those possible modifications) in order to calculate a description length score (and attempt to minimize this score); to better tune the hyperparameters “adjusting the learning rate” based on the error calculated, some embodiments attempt to minimize a description length score that specifies a description length of the trained network (e.g., a number of bits required to describe the network). One possible calculation for such a description length is the number of bits to describe the parameters of the trained network (which would push weight values to 0). However, rather than computing the description length score based on this metric, in some embodiments the description length score calculator 370 uses a measure of the number of bits required to reconstruct the trained network through a prequential hyperparameter tuning technique . The optimization algorithm for the description length score thus seeks to minimize the sum of (i) the bits required to specify the correct output value for each new training input and (ii) the bits required to update the hyperparameters at each iteration ; the disclosure above indicates input indicating feedback to one or more comparisons of a first output and a second output of the one or more neural networks ; see also claims 1 and 10, a memory circuit, and processing unit for executing instructions for: using a first set of inputs to train parameters of the machine-trained (MT) network according to a set of hyperparameters that define aspects of the training by (i ) computing a value of a first loss function based on propagation of the first set of inputs through the MT network i.e., measuring the difference between a first set output generated by the MT network and an expected output for the input, and (ii) modifying the MT network parameters based on gradients of the first loss function with respect to the parameters at the computed value; using a second set of inputs to validate the MT network as trained by the first set of inputs by: propagating the second set of inputs through the MT network with the modified parameters to generate a second set of outputs; and for each input of the second set of inputs, measuring a difference between (i) the output generated by propagating the input through the MT network with the modified parameters and (ii) an expected output for the input; and based on the validation, modifying the hyperparameters for subsequent training of the MT network based on gradients of a description length score with respect to the hyperparameters, wherein the description length score constrains the hyperparameter modification to prevent overfitting of the modified hyperparameters to the second set of inputs by accounting for (i) the difference measurements for each input of the second set of inputs and (ii) the modifications to the hyperparameters ). Teig does not explicitly disclose but Prabhakara discloses the input indicating a selection by one or more users, wherein the user input is based at least in part, on one or more user comparisons between a first output and a second output of the one or more neural networks, presented to the one or more users using a user interfac e ( Prabhakara, col. 3 lines 43-56, model builder 212 determines a best set of hyperparameters by determining/selecting a best model trained and tested using the training data set and the test data set of the historical data. The best set of hyperparameters is used to train a prediction model with the full historical data set. The output of the prediction model is compared to a forecast that is input via forecast interface 204 of interface 202. The comparison is used to identify anomalies and these are provided to the user via feedback module 214 of applications 210 and feedback interface 206 of interface 202. User feedback indicating valid detected anomalies and anomalies not detected is used to retrain the prediction model using model builder 21 2). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to incorporate Prabhakara’s features into Teig’s invention for improving the accuracy of the learning model by using user’s evaluations of the output of the models as data for tuning the model. Regarding claim 22, Teig-Prabhakara discloses the one or more processors of claim 21, further comprising a first learning parameter and a second learning parameter, wherein the first learning parameter comprises a first learning rate, and the second learning parameter comprises a second learning rate that is less than the first learning rate ( Teig, col. 12, line 64 to col. 13, line 12, optimization techniques are used to modify the hyperparameters of machine-trained network. The hyperparameter modifier 375, in concert with the description length score calculator 370, determines the optimal modifications to the hyperparameters 340 at each iteration, and provides these updates to the training system 300. These modifications, for example, might modify the learning rate from one training iteration to another (i.e., to modify the rate at which weight values are changed during backpropagation), increase or decrease regularization factors which tend to push weight values towards 0 in order to reduce overfitting ). The disclosure above indicates modifying at each iteration “the first learning parameter and the second learning parameter” might cause the second learning rate when adjusting the hyperparameters to be less than the first learning rate). Regarding claim 24, Teig-Prabhakara discloses the one or more processors of claim 21, wherein the one or more neural networks include a first set of hyperparameters and a second set of hyperparameters, wherein the first set of hyperparameters includes a first training iteration count and the second set of hyperparameters comprises a second training iteration count that is less than the first training iteration count ( Teig, col. 12, line 64 to col. 13, line 12, optimization techniques are used to modify the hyperparameters of machine-trained network. The hyperparameter modifier 375, in concert with the description length score calculator 370, determines the optimal modifications to the hyperparameters 340 at each iteration, and provides these updates to the training system 300. These modifications, for example, might modify the learning rate from one training iteration to another (i.e., to modify the rate at which weight values are changed during backpropagation), increase or decrease regularization factors which tend to push weight values towards 0 in order to reduce overfitting ). The disclosure above indicates the iteration count when modifying the learning rate might be less than the iteration count when modifying the hyperparameters modifying the at each iteration “the first learning parameter and the second learning parameter” might change the learning rate from one training iteration to another) . Regarding claim 26, Teig-Prabhakara discloses the one or more processors of claim 21, wherein the user input is an input indication from the user interface ( Prabhakara, col. 3 lines 43-56, model builder 212 determines a best set of hyperparameters by determining/selecting a best model trained and tested using the training data set and the test data set of the historical data. The best set of hyperparameters is used to train a prediction model with the full historical data set. The output of the prediction model is compared to a forecast that is input via forecast interface 204 of interface 202. The comparison is used to identify anomalies and these are provided to the user via feedback module 214 of applications 210 and feedback interface 206 of interface 202. User feedback indicating valid detected anomalies and anomalies not detected is used to retrain the prediction model using model builder 21 2). The obviousness argument for combining the references is the same as claim 21. Claims 28-29, 31, 33, 35-37 and 39 comprise limitations substantially the same as claims 21-22, 24, and 26; therefore, they are rejected for the same reasons set forth. Teig-Prabhakara further discloses processing unit, memory circuit with stored-in instructions ( see Teig, fig. 7 ). 6. Claims 23, 25, 27, 30, 32, 34, 38, and 40 are rejected under AIA 35 U.S.C. 103 as being unpatentable over Teig-Prabhakara, as applied to claims 21, 28, and 35 above, in view of Hein et al. (US Publication 2021/0012543, hereinafter Hein). Regarding claim 23, Teig-Prabhakara discloses the one or more processors of claim 21, comprising. Teig-Prabhakara does not explicitly disclose but Hein discloses wherein the one more neural networks are to generate output data including visual artifacts, and wherein the visual artifacts include one or more of geometric aliasing artifacts, rendering noise artifacts, and lighting effect artifacts, and wherein the one or more user comparisons of the first output and the second output is based, at least in part, on the visual artifacts ( Hein, claim 13 discloses a wide variety of artifacts, para’s 0067 and 0072, multiple steps in methods 100 and 200 can be performed using DL networks. In each of these cases, the DL networks can be trained using an appropriate set of training data. In each case, the training data is selected to have input data paired with target data. Each of the DL networks is trained by applying pieces of the input data to the respective DL network to generate an output, and then using a loss function to compare the output from the DL network to the corresponding target data. The DL network is trained by iterative adjusting weighting coefficients of the network to minimize the loss function; in step 254 of process 250, the reconstructed image is applied to an image-domain CNN. The output from the image CNN can be similar to the output from the sinogram CNN, e.g ., the output from the image CNN can include indications of whether an artifact is detected, which artifacts are detected, whether the artifact is attributable to the scanning protocol, whether the artifact is correctable, and whether the artifact is fatal ). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to incorporate Hein’s features into Teig-Prabhakara’s invention for enhancing the output images by incorporating measures for handling artifacts in the training data sets. Regarding claim 25, Teig-Prabhakara discloses the processor of claim 21. Teig-Prabhakara does not explicitly disclose but Hein discloses wherein the user input contains an indication that a completion requirement has been satisfied ( Hein, para. 0109, in operation 485 of process 435 , predefined stopping criteria are used to determine whether the training of the network is complete. For example, the predefined stopping criteria can evaluate whether the new error and/or the total number of iterations performed exceed predefined values. For example, the stopping criteria can be satisfied if either the new error falls below a predefined threshold or if a maximum number of iterations are reached). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to incorporate Hein’s features into Teig-Prabhakara’s invention for enhancing system performance by enabling user to control validation of the operation. Regarding claim 27, Teig-Prabhakara discloses the processor of claim 21, wherein the one or more neural networks contains a first training dataset and a second training dataset, and wherein a learning rate of the second training dataset is adjusted to produce a first microtrained neural network, and the one or more neural networks generate and display a test image from a corresponding training image within the second training dataset using the first microtrained neural network ( Teig, figures 3 and 4, col. 13 through col. 16, initial training input and subsequent training inputs; col. 12, line 64 to col. 13, line 12, hyperparameter modifications might modify the learning rate from one training iteration to another (i.e., to modify the rate at which weight values are changed during backpropagation), increase or decrease regularization factors which tend to push weight values towards 0 in order to reduce overfitting, and as a result produce a micro-trained network that allows adjusting hyperparameters to reduce overfitting; col. 8, lines 7-20, FIG. 3 conceptually illustrates a training system 300 of some embodiments that iteratively adds inputs from a validation set to the training set over the course of multiple training runs. The training system 300 uses a validation system 350 to test the predictivity of the trained network after each iteration and uses a description length score based on (i) potential hyperparameter modifications and (ii) the error generated for validation set inputs when incorporating these potential modifications in order to determine optimal hyperparameter modifications at each iteration; col.38, lines 28-38, the output devices 1445 display images generated by the electronic system ). Teig-Prabhakara does not explicitly disclose but Hein discloses the one or more neural networks generates output data including visual artifacts, wherein the visual artifacts are reduced within the test image relative to a second test image generated by a neural network of the corresponding training image ( Hein, claim 13 discloses a wide variety of artifacts in the input images, and generating output images to reduce artifacts ). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to incorporate Hein’s features into Teig-Prabhakara’s invention for improving system performance and capability by providing operation for handling artifacts. Regarding claims 30, 32, 34, 38, and 40, these claims comprise limitations substantially the same as claims 23, 25, and 27; therefore, they are rejected for the same reasons set forth. Conclusion 07-40 AIA 7. 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. 8. Any inquiry concerning this communication or earlier communications from the examiner should be directed to LOI H TRAN whose telephone number is (571)270-5645. The examiner can normally be reached 8:00AM-5:00PM PST FIRST FRIDAY OF BIWEEK OFF. 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, THAI TRAN can be reached at 571-272-7382. 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. /LOI H TRAN/Primary Examiner, Art Unit 2484 Application/Control Number: 16/818,266 Page 2 Art Unit: 2484 Application/Control Number: 16/818,266 Page 3 Art Unit: 2484 Application/Control Number: 16/818,266 Page 4 Art Unit: 2484 Application/Control Number: 16/818,266 Page 5 Art Unit: 2484 Application/Control Number: 16/818,266 Page 6 Art Unit: 2484 Application/Control Number: 16/818,266 Page 7 Art Unit: 2484 Application/Control Number: 16/818,266 Page 8 Art Unit: 2484 Application/Control Number: 16/818,266 Page 9 Art Unit: 2484
Prosecution Insights