DETAILED ACTION
This Office Action is sent in response to the Applicant’s Communication received on 02/18/2026 for application number 17/592,186. The Office hereby acknowledges receipt of the following and placed of record in file: Specification, Drawings, Abstract, Oath/Declaration, IDS, and Claims.
1, 4-6, 9, 13 and 16 are amended.
Claims 7 and 8 are canceled
Claims 21 and 22 are new.
Claims 1-6, 9-22 are pending.
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
35 USC 101
In summary, on pages 17-19 of the remarks section, Applicant argues that the features of amended claim 1 provide a particular technical solution to the technological problems of inaccuracy and overfitting associated with conventional array-completion methods used on an incomplete array by using influenced-based data augmentation and training for an array completion model. Further, the features of amended claim 1 improve the functionality of the computer by enabling array-completion models to more effectively handle incomplete arrays (e.g., sparse arrays) and train array-completion models using incomplete arrays compared to conventional array-completion techniques. As discussed in the specification, embodiments generate new data points for cells by sampling entities proportional to their entity-importance metrics, values of the cells are predicted via a trained neural tensor completion model (e.g., either the first completion model or the second completion model), and the influence-based sampling of entities is employed to generate augmented data points for the cells by using important entities (e.g., as weighted by their entity-importance metrics in the sampling process), which can lead to higher test prediction accuracy than conventional tensor completion methods. Amended claim 1 specifically recites receiving an incomplete array comprising a subset of cells that are empty, store an initialized zero value, or store a null value, which links amended claim 1 to the technological problems associated with incomplete arrays. The identification of the first cell based on the weighted sampling and the generation of the augmentation array comprising the first cell storing the predicted value as claimed corresponds to the features discussed in paragraph [0023] of the present application that produce the improved test prediction accuracy. Amended claim 1 recites that an augmented array is generated based on a union of the incomplete array and the augmentation array that comprises the first cell storing the predicted value and that one or more parameters of an array-completion model are adapted based on the augmented array. Generating the augmented array and adapting the parameter(s) of the array-completion model based on the augmented array as claimed corresponds to the features of data augmentation during training discussed in paragraph [0026] of the present application that produce the benefit of enhanced prediction accuracy.
Furthermore, in summary of pages 20 and 21 of the remarks section, Applicant argues that amended claim 9 also recites "generating an augmentation tensor comprising a predicted value for the first cell, the predicted value for the first cell inferred by a second tensor completion model based at least on one or more other cells of the incomplete tensor from the first subset of cells." As discussed in the specification, conventional methods (e.g., tensor factorization) may suffer from overfitting when the input tensor is very sparse. The use of a second tensor-completion model to infer the predicted value, as recited in claim 9, provides a further technical improvement by increasing the generalization capability of a downstream model and reducing overfitting. In addition to the improved accuracy discussed above, the combined features of claim 9 provide an additional technical improvement by employing a second model to generate heterogeneous augmented data that increases generalization capability and avoids the overfitting problems of conventional methods.
In light of the newly amended claim limitations, the Examiner finds the Applicant’s argument persuasive. Therefore, the 35 USC 101 rejection is withdrawn.
35 USC 103
On pages 23 and 24 of the remarks section, that nothing in the asserted combination of Kenji and Zhu teaches or suggests "receiving an incomplete array having a multi-dimensional data structure comprising an indexed set of cells associated with a set of entities, wherein the set of cells comprises a first subset of cells that store a relevant value and a second subset of cells that are empty, store an initialized zero value, or store a null value" as recited in amended claim 1. Nothing in the cited portions of Kenji teaches, or even suggests, that cells in the dataset DS 1 are empty, store an initialized zero value, or store a null value as claimed.
The Examiner respectfully disagrees. The limitation “receiving an incomplete array having a multi-dimensional data structure comprising an indexed set of cells associated with a set of entities” is taught by Kenji in paragraph 0014, “In FIG. 1, each square in the data set DS1 (an indexed set of cells) represents data (image) (associated with a set of entities), and the data set DS1 includes a large number of data (images) (having a multi-dimensional data),” paragraph 0016, “a dataset DS1 in which each data (image) is associated with a correct answer label indicating the presence or absence of a smile”, paragraph 0028, “Then, the data adjustment apparatus 100 adjusts the data set DS1 by adding data corresponding to the data DT33 to the data set DS1”, and paragraph 0060, “Each image is assigned a label, i.e., an image is associated with a correct label. For example, if there is a set (dataset) of n images (n is any natural number) and labels, each labeled image z”. Also see an incomplete array Figure 1. Applicant’s arguments with respect to the limitation “wherein the set of cells comprises a first subset of cells that store a relevant value and a second subset of cells that are empty, store an initialized zero value, or store a null value” have been considered but are moot because the new ground of rejection relies on newly introduced reference Tang.
On pages 24 and 25 of the remarks section, Applicant argues that nothing in the asserted combination of Kenji and Zhu teaches or suggests "generating an augmentation array comprising a predicted value for the first cell, the predicted value inferred by the array-completion model or a second array-completion model based at least on one or more other cells of the incomplete array from the first subset of cells" as recited in amended claim 1. By the language of amended claim 1, the "first cell" is identified as an incomplete cell, "from the second subset of cells," and the second subset of cells "are empty, store an initialized zero value, or store a null value." Nothing in the cited portions of Kenji teaches, or even suggests, generating an augmentation array comprising a predicted value for a particular incomplete cell inferred using an array-completion model.
The Examiner respectfully disagrees. Kenji teaches “the first cell” in paragraph 0021, “the data adjusting device 100 measures the degree of influence that the data DT14 in the data set DS1”. However, Applicant’s arguments with respect to the limitation “generating an augmentation array comprising a predicted value for cell, the predicted value inferred by array-completion model or a second array-completion model based at least on one or more other cells of the incomplete array from the first subset of cells” have been considered but are moot because the new ground of rejection relies on newly introduced reference Tang.
On pages 25 to 26 of the remarks section, Applicant argues that nothing in the asserted combination of Kenji and Zhu teaches or suggests "generating an augmented array having the multi-dimensional data structure, and that includes a union of the incomplete array with the augmentation array comprising the first cell storing the predicted value" as recited in amended claim 1. Nothing in the cited portions of Kenji teaches, or even suggests, that the adjusted dataset has the same multi-dimensional structure as the original dataset prior to adjustment. The cited portions of Kenji are silent with respect to a particular structure of the dataset DS1 and generically describe that low contribution data is removed and data similar to high contribution data is added. See, for example, paragraph [0039] of Kenji. Further, nothing in the cited portions of Kenji teaches, or even suggests, a union of the incomplete array with the augmentation array comprising a previously incomplete cell that stored a predicted value inferred by an array-completion model. Instead, Kenji describes generating the adjusted dataset DS1 by removing low contribution data from the original dataset and adding new data that is similar to high contribution data in the original dataset DS1.
The Examiner respectfully disagrees. Kenji does indeed teach “generating an augmented array having the multi-dimensional data structure, and that includes a union of the incomplete array with the augmentation array” in paragraph 0028, “Therefore, as indicated by the discrimination result DR2, the data adjustment apparatus 100 determines that the data DT33 has a high contribution to the learning of the model M1, and determines to add data corresponding to the data DT33 to the data set DS1. Then, the data adjustment apparatus 100 adjusts the data set DS1 by adding data corresponding to the data DT33 to the data set DS1 (includes a union of the incomplete array with the augmentation array). In this way, the data adjustment device 100 updates the data set DS1”, paragraph 0034, “The data adjusting device 100, which has acquired the provision data from the terminal device 10, adds the acquired provision data to the data set DS1 (generating an augmented array) (step S9). As a result, the data adjusting device 100 adds data similar to the data DT33 with a high contribution rate to the data set DS1”. Figure 1 of Kenji discloses the multi-dimensional data structure. However, Applicant’s arguments with respect to the limitation “augmentation array comprising cell storing the predicted value” have been considered but are moot because the new ground of rejection relies on newly introduced reference Tang.
On page 27 of the remarks section, Applicant argues that that nothing in the asserted combination of Kenji and Zhu teaches or suggests "adapting one or more parameters of the array-completion model or the second array-completion model based at least on the augmented array" as recited in amended claim 1. Rather than adapting one or more parameters of an array-completion model as claimed, the cited portions of Kenji describe training a neural network (model M1) that performs image recognition and specifically discusses that the model may perform smile detection, object recognition, emotion detection, or other classification operations.
The Examiner respectfully disagrees. Kenji does indeed teach “adapting one or more parameters of model or second model” in paragraph 0016, “The data adjusting apparatus 100 may use various functions as the loss function as long as the influence of each piece of data can be measured in the measurement process; Para 0017, in the backpropagation method, a loss function is used to indicate how far the value of the output layer of the neural network (model or second model) is from the correct state (correct label), and the weights and biases are updated (adapting one or more parameters) using the steepest descent method or the like so as to minimize the loss function… the data adjustment device 100 provides input values (data) to a neural network (model M1), which calculates a predicted value based on the input values, and compares the predicted value with teacher data (correct label) to evaluate the error”. However, Applicant’s arguments with respect to the limitation “model being array-completion model or second array-completion model based at least on the augmented array” have been considered but are moot because the new ground of rejection relies on newly introduced reference Tang.
Therefore, the 35 USC 103 rejection is maintained.
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.
Claims 1, 9-11, 14, 16, 20 and 22 are rejected under 35 U.S.C. 103 as being unpatentable over Kenji (WO 2021200392 A1, see attached documents of description and drawings, hereinafter Kenji) in view of Zhu et al. (U.S. 20220383185 A1), hereinafter Zhu, and Tang et al. (CN113687297A, see attached translation), hereinafter Tang.
Regarding claim 1, Kenji teaches
A non-transitory computer-readable storage medium (Para 0232, the HDD 1400) having instructions stored thereon (Para 0232, a program), which, when executed by a processing device (Para 0232, The CPU 1100) [Para 0232, The CPU 1100 operates based on a program stored in the ROM 1300 or the HDD 1400, and controls each unit.], cause the processing device to perform operations comprising:
receiving an incomplete array having a multi-dimensional data structure (Fig. 1, DS1) comprising an indexed set of cells (Fig. 1, DT14 and DT33) associated with a set of entities (Para 0014, images) [Para 0014, DS1 includes a large number of data (images)) [Para 0014, In FIG. 1, each square in the data set DS1 represents data (image), and the data set DS1 includes a large number of data (images); Para 0016, a dataset DS1 in which each data (image) is associated with a correct answer label indicating the presence or absence of a smile; Para 0028, Then, the data adjustment apparatus 100 adjusts the data set DS1 by adding data corresponding to the data DT33 to the data set DS1; Para 0060, Each image is assigned a label, i.e., an image is associated with a correct label. For example, if there is a set (dataset) of n images (n is any natural number) and labels, each labeled image z];
determining influence metrics of cells from the incomplete array based on applying an influence function [Fig. 1, MM1, influence function] to estimate an influence of the cells on minimizing a loss signal generated from training an array-completion (Para 0016, data adjusting apparatus) model using the incomplete array (Para 0018, data set DS1) [Para 0012, the data adjustment device 100 performs a process of adjusting data from a dataset used to train a deep neural network (DNN); Para 0016, The data adjusting apparatus 100 may use various functions as the loss function as long as the influence of each piece of data can be measured in the measurement process; Para 0017, in the backpropagation method, a loss function is used to indicate how far the value of the output layer of the neural network is from the correct state (correct label), and the weights and biases are updated using the steepest descent method or the like so as to minimize the loss function… the data adjustment device 100 provides input values (data) to a neural network (model M1), which calculates a predicted value based on the input values, and compares the predicted value with teacher data (correct label) to evaluate the error; Para 0018, Then, the data adjusting device 100 measures the degree of influence that each data in the data set DS1 has on the learning of the model M1. The data adjustment device 100 uses a method for measuring the degree of influence (measurement method MM1) to measure the degree of influence that each data item in the data set DS1 has on the learning of the model M1];
generating an entity-importance metric (Para 0018, degree of contribution, influence level IV14 and influence level IV33) for each entity of the set of entities based on an aggregation (Para 0030, The terminal device 10 collects data similar to the data DT33 as data to be provided to the data adjustment device 100) of the influence metrics of the cells (Para 0018, the degree of influence that each data item in the data set DS1) associated with the entity [Para 0018, The data adjustment device 100 uses a method for measuring the degree of influence (measurement method MM1) to measure the degree of influence that each data item in the data set DS1 has on the learning of the model M1. The degree of influence here indicates that the greater the value, the greater the degree to which the data contributed to the learning of the model M1 (degree of contribution). A larger influence value, i.e., a higher influence, indicates a greater contribution to improving the classification accuracy of model M1. In this way, the higher the influence level, the more necessary the data is for training the model M1. For example, a higher influence indicates that the data is more beneficial for learning the model M1; Para 0019, Furthermore, the smaller the influence value, the lower the degree to which the data contributed to the learning of the model M1 (degree of contribution); Para 0021, In FIG. 1, the data adjusting device 100 measures the degree of influence that the data DT14 in the data set DS1 has on the learning of the model M1 (step S2). As shown in the measurement result RS1, the data adjusting device 100 measures the degree of influence that the data DT14 has on the learning of the model M1 as an influence IV14. The influence level IV14 is assumed to be a specific value (for example, 0.2); Para 0030, The terminal device 10 collects data similar to the data DT33 as data to be provided to the data adjustment device 100]
identifying a first cell from the array as an incomplete cell [Para 0021, the data adjusting device 100 measures the degree of influence that the data DT14 in the data set DS1] based on a sampling of the set of entities that is weighted (Para 0017, successively correcting the values of the connection weights) [Para 0017, the data adjusting device 100 trains the model M1 by updating parameters such as weights and biases so that the output layer has correct values for the input data. For example, in the backpropagation method, a loss function is used to indicate how far the value of the output layer of the neural network is from the correct state (correct label), and the weights and biases are updated using the steepest descent method or the like so as to minimize the loss function. For example, the data adjustment device 100 provides input values (data) to a neural network (model M1), which calculates a predicted value based on the input values, and compares the predicted value with teacher data (correct answer labels) to evaluate the error. Then, the data adjusting device 100 executes learning and construction of the model M1 by successively correcting the values of the connection weights (synaptic coefficients) in the neural network (model M1) based on the obtained errors];
generating an augmented array having the multi-dimensional data structure, and that includes a union of the incomplete array with the augmentation array [Para 0028, Therefore, as indicated by the discrimination result DR2, the data adjustment apparatus 100 determines that the data DT33 has a high contribution to the learning of the model M1, and determines to add data corresponding to the data DT33 to the data set DS1. Then, the data adjustment apparatus 100 adjusts the data set DS1 by adding data corresponding to the data DT33 to the data set DS1. In this way, the data adjustment device 100 updates the data set DS1; Para 0033, Then, the terminal device 10 provides the provision data to the data adjusting device 100 (step S8). The terminal device 10 transmits data similar to the collected data DT33 to the data adjusting device 100 as data to be provided; Para 0034, The data adjusting device 100, which has acquired the provision data from the terminal device 10, adds the acquired provision data to the data set DS1 (step S9). As a result, the data adjusting device 100 adds data similar to the data DT33 with a high contribution rate to the data set DS1; Note: instant application specification refers to an incomplete array as being an array containing non-relevant values in Para 0004]; and
adapting one or more parameters of model or second model [Para 0016, The data adjusting apparatus 100 may use various functions as the loss function as long as the influence of each piece of data can be measured in the measurement process; Para 0017, in the backpropagation method, a loss function is used to indicate how far the value of the output layer of the neural network is from the correct state (correct label), and the weights and biases are updated using the steepest descent method or the like so as to minimize the loss function… the data adjustment device 100 provides input values (data) to a neural network (model M1), which calculates a predicted value based on the input values, and compares the predicted value with teacher data (correct label) to evaluate the error].
Kenji does not teach wherein the set of cells comprises a first subset of cells that store a relevant value and a second subset of cells that are empty, store an initialized zero value, or store a null value; set of entities that is weighted proportional to the entity-importance metric for each entity of the set of entities; generating an augmentation array comprising a predicted value for the first cell, the predicted value inferred by the array-completion model or a second array- completion model based at least on one or more other cells of the incomplete array from the first subset of cells; augmentation array comprising the first cell storing the predicted value; and model being array-completion model or second array-completion model based at least on the augmented array.
Zhu teaches,
set of entities (Para 0008, training examples, zi) that is weighted proportional to the entity-importance metric (Para 0049, i-specific weighing parameter
PNG
media_image1.png
55
55
media_image1.png
Greyscale
) for each entity of the set of entities [Para 0028, calculating importance scores for the training examples; ranking the importance scores (e.g., TracInF(z,z′)) to create a ranked list of the training examples;… For instance, the importance scores TracInF(z,z′) can be calculated as:
PNG
media_image2.png
296
920
media_image2.png
Greyscale
]
Zhu is analogous to the claimed invention as they both relate to the implementation of influence function in neural network predictions. 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 Kenji’s teachings to incorporate the teachings of Zhu and provide set of entities that is weighted by the entity-importance metric for each entity of the set of entities [Zhu, para 0024] to enhance the interpretability of the explanations (i.e., training examples).
Kenji-Zhu teach the above limitations of claim 1 including the set of cells (Kenji, Fig. 1 and para 0014) and the first cell (Kenji, para 0021).
Kenji-Zhu do not teach wherein cells comprises a first subset of cells that store a relevant value and a second subset of cells that are empty, store an initialized zero value, or store a null value; generating an augmentation array comprising a predicted value for cell, the predicted value inferred by the array-completion model or a second array- completion model based at least on one or more other cells of the incomplete array from the first subset of cells; augmentation array comprising cell storing the predicted value; and model being array-completion model or second array-completion model based at least on the augmented array.
Tang teaches,
wherein cells comprises a first subset of cells that store a relevant value and a second subset of cells that are empty, store an initialized zero value, or store a null value [Para 0011, Step 2: Utilizing the low-rank property of the received signal tensor model, the received signal tensor model constructed in Step 1 is filled with missing data using a matrix factorization based tensor completion method to obtain the completed tensor model];
generating an augmentation array comprising a predicted value for cell, the predicted value inferred by array-completion model (Para 0064, DOA estimation algorithms) [Para 0012, Step 3: Expand the completed tensor model to obtain the completed received signal matrix. Use the ESPRIT algorithm for acoustic vector sensor arrays to estimate the source direction of arrival (DOA) of the completed received signal matrix; Para 0064, This invention utilizes a matrix decomposition-based tensor completion algorithm to fold and unfold the received signal data of any acoustic vector sensor array, completing the missing data. This combines tensor completion with traditional DOA estimation algorithms, ultimately achieving accurate DOA estimation even when array element damage results in partial loss of received signal data.] or a second array-completion model based at least on one or more other cells of the incomplete array from the first subset of cells (alternate);
augmentation array comprising cell storing the predicted value [Para 0012, Step 3: Expand the completed tensor model to obtain the completed received signal matrix. Use the ESPRIT algorithm for acoustic vector sensor arrays to estimate the source direction of arrival (DOA) of the completed received signal matrix; Para 0064, This invention utilizes a matrix decomposition-based tensor completion algorithm to fold and unfold the received signal data of any acoustic vector sensor array, completing the missing data. This combines tensor completion with traditional DOA estimation algorithms, ultimately achieving accurate DOA estimation even when array element damage results in partial loss of received signal data; Para 0118, The received signal with missing data is filled in using the matrix factorization-based tensor completion algorithm (MFTC) to obtain the complete tensor data. Expanding this tensor yields the complete received signal matrix];
and model being array-completion model [Para 0064, This invention utilizes a matrix decomposition-based tensor completion algorithm to fold and unfold the received signal data of any acoustic vector sensor array, completing the missing data. This combines tensor completion with traditional DOA estimation algorithms, ultimately achieving accurate DOA estimation even when array element damage results in partial loss of received signal data] or second array-completion model based at least on the augmented array (alternate).
Tang is analogous to the claimed invention as they both relate to data completion. 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 Kenji’s teachings to incorporate the teachings of Tang and provide using an array-completion model to generate an augmentation array in order to [Tang, para 0004] effectively improve estimation algorithms.
Regarding claim 9, Kenji teaches,
A method comprising: receiving an incomplete tensor having a multi-dimensional data structure (Fig. 1, DS1) comprising an indexed set of cells (Fig. 1, DT14 and DT33) associated with a set of entities (Para 0014, images) [Para 0014, In FIG. 1, each square in the data set DS1 represents data (image), and the data set DS1 includes a large number of data (images); Para 0016, a dataset DS1 in which each data (image) is associated with a correct answer label indicating the presence or absence of a smile; Para 0028, Then, the data adjustment apparatus 100 adjusts the data set DS1 by adding data corresponding to the data DT33 to the data set DS1],
the indexed set of cells including a first subset of cells (Para 0039, DT33) that store a relevant value (Para 0039, data with high contribution) and a second subset of cells (Para 0021, DT14) that store a value (Para 0039, data with low contribution) [Para 0021, In FIG. 1, the data adjusting device 100 measures the degree of influence that the data DT14 in the data set DS1 has on the learning of the model M1; Para 0025, Also, in FIG. 1, the data adjusting device 100 measures the degree of influence that the data DT33 in the data set DS1 has on the learning of the model M1; Para 0039, the data adjustment device 100 re-trains the model M1 using an adjusted dataset DS1 in which data with low contribution, such as data DT14, is excluded and data similar to data with high contribution, such as data DT33, is added]; and
generating an augmented tensor having the multi-dimensional data structure (See Fig. 1; Para 0039, adjusted dataset DS1) by:
determining influence metrics (Para 0018, degree of influence) of cells from the incomplete tensor (Para 0016, influence of each piece of data can be measured in the measurement process) based on applying an influence function [Fig. 1, MM1, influence function] to estimate an influence of the cells on minimizing a loss signal (Para 0017, the data adjustment device 100 provides input values (data) to a neural network (model M1), which calculates a predicted value based on the input values, and compares the predicted value with teacher data (correct label) to evaluate the error) generated from training a first tensor-completion model (Para 0012, data adjustment device) using the incomplete tensor (Para 0012, the data adjustment device 100 performs a process of adjusting data from a dataset used to train a deep neural network) [Para 0012, the data adjustment device 100 performs a process of adjusting data from a dataset used to train a deep neural network (DNN); Para 0014, In FIG. 1, each square in the data set DS1 represents data (image), and the data set DS1 includes a large number of data (images); Para 0016, The data adjusting apparatus 100 may use various functions as the loss function as long as the influence of each piece of data can be measured in the measurement process; Para 0017, in the backpropagation method, a loss function is used to indicate how far the value of the output layer of the neural network is from the correct state (correct label), and the weights and biases are updated using the steepest descent method or the like so as to minimize the loss function… the data adjustment device 100 provides input values (data) to a neural network (model M1), which calculates a predicted value based on the input values, and compares the predicted value with teacher data (correct label) to evaluate the error; Para 0018, Then, the data adjusting device 100 measures the degree of influence that each data in the data set DS1 has on the learning of the model M1. The data adjustment device 100 uses a method for measuring the degree of influence (measurement method MM1) to measure the degree of influence that each data item in the data set DS1 has on the learning of the model M1];
selecting a first cell from the second subset of cells of the incomplete tensor (Para 0039, removes data with a low contribution) based on a sampling (Para 0018, measuring the degree of influence) determined from training the first tensor-completion model (Para 0039, re-trains the model M1) using the incomplete tensor, the entity-importance metric for each respective entity of the set of entities generated (Para 0018, degree of contribution, influence level IV14 and influence level IV33) based on an aggregation (Para 0030, The terminal device 10 collects data similar to the data DT33 as data to be provided to the data adjustment device 100) of the influence metrics of cells (Para 0018, the degree of influence that each data item in the data set DS1) associated with the respective entity [Para 0018, The data adjustment device 100 uses a method for measuring the degree of influence (measurement method MM1) to measure the degree of influence that each data item in the data set DS1 has on the learning of the model M1. The degree of influence here indicates that the greater the value, the greater the degree to which the data contributed to the learning of the model M1 (degree of contribution). A larger influence value, i.e., a higher influence, indicates a greater contribution to improving the classification accuracy of model M1. In this way, the higher the influence level, the more necessary the data is for training the model M1. For example, a higher influence indicates that the data is more beneficial for learning the model M1; Para 0038, For example, the data adjusting device 100 measures the degree of influence for all data in the data set DS1. Then, the data adjusting device 100 removes data with a low contribution rate from the data set DS1. In addition, the data adjusting device 100 adds data similar to the data with a high contribution to the data set DS1; Para 0039, the data adjustment device 100 re-trains the model M1 using an adjusted dataset DS1 in which data with low contribution, such as data DT14, is excluded and data similar to data with high contribution, such as data DT33, is added]; and
wherein the augmented tensor (Para 0039, adjusted dataset DS1) comprises a union of the incomplete tensor with the augmentation tensor comprising the first cell storing (Para 0039, is added) the predicted value (Para 0039, data DT33) [Para 0028, Therefore, as indicated by the discrimination result DR2, the data adjustment apparatus 100 determines that the data DT33 has a high contribution to the learning of the model M1, and determines to add data corresponding to the data DT33 to the data set DS1. Then, the data adjustment apparatus 100 adjusts the data set DS1 by adding data corresponding to the data DT33 to the data set DS1. In this way, the data adjustment device 100 updates the data set DS1; Para 0039, data adjustment device 100 re-trains the model M1 using an adjusted dataset DS1 in which data with low contribution, such as data DT14, is excluded and data similar to data with high contribution, such as data DT33, is added] ; and
adapting one or more parameters of model or second model [Para 0016, The data adjusting apparatus 100 may use various functions as the loss function as long as the influence of each piece of data can be measured in the measurement process; Para 0017, in the backpropagation method, a loss function is used to indicate how far the value of the output layer of the neural network is from the correct state (correct label), and the weights and biases are updated using the steepest descent method or the like so as to minimize the loss function… the data adjustment device 100 provides input values (data) to a neural network (model M1), which calculates a predicted value based on the input values, and compares the predicted value with teacher data (correct label) to evaluate the error].
Kenji does not teach wherein cells that are empty store an initialized zero value, or store a null value; a sampling of the set of entities that is weighted proportional to the entity-importance metrics; generating an augmentation tensor comprising a predicted value for cell, the predicted value inferred by a second tensor-completion model based at least on one or more other cells of the incomplete array from the first subset of cells; and model being tensor-completion model or second tensor-completion model based at least on the augmented tensor.
Zhu teaches,
a sampling [Para 0008, using training examples from randomly selected batches of the training data D] of the set of entities (Para 0008, training examples, zi) that is weighted proportional to the entity-importance metric (Para 0049, i-specific weighing parameter
PNG
media_image1.png
55
55
media_image1.png
Greyscale
) [Para 0028, calculating importance scores for the training examples; ranking the importance scores (e.g., TracInF(z,z′)) to create a ranked list of the training examples;… For instance, the importance scores TracInF(z,z′) can be calculated as:
PNG
media_image2.png
296
920
media_image2.png
Greyscale
]
Zhu is analogous to the claimed invention as they both relate to the implementation of influence function in neural network predictions. 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 Kenji’s teachings to incorporate the teachings of Zhu and provide set of entities that is weighted by the entity-importance metric for each entity of the set of entities [Zhu, para 0024] to enhance the interpretability of the explanations (i.e., training examples).
Kenji-Zhu teach the above limitations of claim 9 including the first cell (Kenji, Para 0039) and the first subset of cells (Kenji, Fig. 1 and para 0014).
Kenji-Zhu do not teach wherein cells that are empty store an initialized zero value, or store a null value; generating an augmentation tensor comprising a predicted value for cell, the predicted value for the cell inferred by a second tensor-completion model based at least on one or more other cells of the incomplete tensor; and model being tensor-completion model or second tensor-completion model based at least on the augmented tensor.
Tang teaches,
wherein cells that are empty store an initialized zero value, or store a null value [Para 0011, Step 2: Utilizing the low-rank property of the received signal tensor model, the received signal tensor model constructed in Step 1 is filled with missing data using a matrix factorization based tensor completion method to obtain the completed tensor model];
generating an augmentation tensor comprising a predicted value for cell, the predicted value inferred by a second tensor-completion model based at least on one or more other cells of the incomplete array from the first subset of cells [Para 0012, Step 3: Expand the completed tensor model to obtain the completed received signal matrix. Use the ESPRIT algorithm for acoustic vector sensor arrays to estimate the source direction of arrival (DOA) of the completed received signal matrix; Para 0064, This invention utilizes a matrix decomposition-based tensor completion algorithm to fold and unfold the received signal data of any acoustic vector sensor array, completing the missing data. This combines tensor completion with traditional DOA estimation algorithms, ultimately achieving accurate DOA estimation even when array element damage results in partial loss of received signal data; Para 0118, The received signal with missing data is filled in using the matrix factorization-based tensor completion algorithm (MFTC) to obtain the complete tensor data. Expanding this tensor yields the complete received signal matrix];
and model being tensor-completion model [Para 0064, This invention utilizes a matrix decomposition-based tensor completion algorithm to fold and unfold the received signal data of any acoustic vector sensor array, completing the missing data. This combines tensor completion with traditional DOA estimation algorithms, ultimately achieving accurate DOA estimation even when array element damage results in partial loss of received signal data] or second tensor-completion model based at least on the augmented tensor (alternate).
Regarding claim 10, Kenji-Zhu-Tang teaches all the limitations of claim 9.
Kenji further teaches,
wherein generating the augmented tensor further comprises: calculating an entity-importance metric for each entity of the set of entities (Para 0038, measures the degree of influence for all data in the data set DS1) based on the training of the first tensor-completion model (Para 0039, re-trains the model M1 using an adjusted dataset DS1) [Para 0038, For example, the data adjusting device 100 measures the degree of influence for all data in the data set DS1. Then, the data adjusting device 100 removes data with a low contribution rate from the data set DS1. In addition, the data adjusting device 100 adds data similar to the data with a high contribution to the data set DS1; Para 0039, the data adjustment device 100 re-trains the model M1 using an adjusted dataset DS1].
Regarding claim 11, Kenji-Zhu-Tang teaches all the limitations of claims 10.
Kenji further teaches,
wherein generating the augmented tensor further comprises: accessing the loss signal (Fig. 1, output of MM1) that was generated during the training (Para 0016, learning process) of the first tensor-completion model (Para 0016, M1) using a set of training cells from the incomplete tensor (Para 0016, data set DS1) [Para 0016, The data adjustment apparatus 100 performs a learning process using the data set DS1 so as to minimize a set loss function, and learns the model M1. The data adjusting apparatus 100 may use various functions as the loss function as long as the influence of each piece of data can be measured in the measurement process];
calculating, based on the loss signal (Fig. 1, output of MM1), a cell-importance metric (Fig. 1, IV33 and IV14) for each training cell (Para 0016, each piece of data) of the set of training cells from the incomplete tensor (Para 0020, the entire data set DS1) [Para 0016, The data adjusting apparatus 100 may use various functions as the loss function as long as the influence of each piece of data can be measured in the measurement; Para 0020, For example, the data adjustment apparatus 100 may measure the difference between the loss in learning when the entire data set DS1 is used and the loss in learning when data X is removed from data set DS1 as the influence of data X; Para 0021, As shown in the measurement result RS1, the data adjusting device 100 measures the degree of influence that the data DT14 has on the learning of the model M1 as an influence IV14. The influence level IV14 is assumed to be a specific value (for example, 0.2); Para 0025, As shown in the measurement result RS2, the data adjusting device 100 measures the degree of influence that the data DT33 has on the learning of the model M1 as an influence IV33. The influence level IV33 is assumed to be a specific value (for example, 0.7)];
and wherein, for each entity of the set of entities [Para 0038, pieces of data in the data set DS1], the entity-importance metric for the entity is determined [Para 0018, degree of contribution, influence level IV14 and influence level IV33] based on a combination of the cell-importance metrics [Para 0038, device 100 executes similar processing for all pieces of data in the data set DS1] that corresponds to the entity (Para 0038, data in the data set DS1) [Para 0021, As shown in the measurement result RS1, the data adjusting device 100 measures the degree of influence that the data DT14 has on the learning of the model M1 as an influence IV14; Para 0025, As shown in the measurement result RS2, the data adjusting device 100 measures the degree of influence that the data DT33 has on the learning of the model M1 as an influence IV33.) that corresponds to the entity; Para 0038, Although FIG. 1 shows processing for only two pieces of data, DT14 and DT33, for the sake of explanation, the data adjustment device 100 executes similar processing for all pieces of data in the data set DS1].
Regarding claim 14, Kenji-Zhu-Tang teaches all the limitations of claim 9 including the first machine learning model.
Kenji further teaches
wherein the entity-importance metrics (Para 0018, degree of contribution, influence level IV14 and influence level IV33) are determined using the influence metrics (Para 0018, the degree of influence that each data item in the data set DS1) [Para 0018, The data adjustment device 100 uses a method for measuring the degree of influence (measurement method MM1) to measure the degree of influence that each data item in the data set DS1 has on the learning of the model M1. The degree of influence here indicates that the greater the value, the greater the degree to which the data contributed to the learning of the model M1 (degree of contribution). A larger influence value, i.e., a higher influence, indicates a greater contribution to improving the classification accuracy of model M1. In this way, the higher the influence level, the more necessary the data is for training the model M1. For example, a higher influence indicates that the data is more beneficial for learning the model M1; Para 0019, Furthermore, the smaller the influence value, the lower the degree to which the data contributed to the learning of the model M1 (degree of contribution); Para 0021, In FIG. 1, the data adjusting device 100 measures the degree of influence that the data DT14 in the data set DS1 has on the learning of the model M1 (step S2). As shown in the measurement result RS1, the data adjusting device 100 measures the degree of influence that the data DT14 has on the learning of the model M1 as an influence IV14. The influence level IV14 is assumed to be a specific value (for example, 0.2);]
Kenji does not teach for each entity of the set of entities, determining an entity-embedding; for each training cell of a set of training cells, generating a cell-embedding based on a combination of the entity-embeddings for a subset of the set of entities that is associated with the training cell; determining estimated values for a set of test cells based on the cell- embedding for each of the training cells of the set of training cells; generating the loss signal based on the estimated values for the set of test cells; and for each training cell of the set of training cells, determining a respective influence metric by employing an influence function that is based on the loss signal.
Zhu further teaches,
for each entity of the set of entities (Para 0033, training span), determining an entity-embedding [Para 0033, According to an exemplary embodiment, the training span is first defined from token i to token j to be xij, and a sequence with xij masked is: x−ij=[x0, . . . ,xi-1,[MASK], . . . ,[MASK],xj+1, . . . ], and its corresponding training data is z−ij];
for each training cell of a set of training cells, generating a cell-embedding (Para 0033, training data is z−ij) based on a combination of the entity-embeddings for a subset of the set of entities that is associated with the training cell [Para 0033, According to an exemplary embodiment, the training span is first defined from token i to token j to be xij, and a sequence with xij masked is: x−ij=[x0, . . . ,xi-1,[MASK], . . . ,[MASK],xj+1, . . . ], and its corresponding training data is z−ij];
determining estimated values [Para 0033, empirical-risk-estimated parameter {circumflex over (θ)} obtained from D.sup.train] for a set of test cells [Para 0037, IF.sup.+ measures the influence of a training span on an entire test sequence…Similarly, the influence of a training span to a test span x.sub.kl′ is also measured] based on the cell- embedding for each of the training cells of the set of training cells; generating (Para 0052, empirical-risk-estimated parameter calculation) a loss signal (Para 0052, loss difference) based on the estimated values for the set of test cells [Para 0052, the logit difference is used as an importance score based on the empirical-risk-estimated parameter {circumflex over (θ)} obtained from D.sup.train as: imp(x.sub.ij|z,{circumflex over (θ)})=logit.sub.y(x;{circumflex over (θ)})−logit.sub.y(x.sub.−ij;{circumflex over (θ)}), wherein every term in the right hand side (RHS) is the logit output evaluated at a model prediction y from model {circumflex over (θ)} to be explained right before applying the SoftMax function, which is a commonly used machine linear function to map input into a probability distribution. This equation indicates how important a training span is, and is equivalent to the loss difference: imp(z.sub.ij|z,θ)=custom-character(z.sub.−ij;{circumflex over (θ)})−custom-character(z;{circumflex over (θ)}) (3)
when the cross-entropy loss custom-character(z;θ)=−Σ.sub.y.sub.icustom-character(y=y.sub.i)logit.sub.y.sub.i(x;θ) is applied.];
and for each training cell of the set of training cells, determining a respective influence metric by employing an influence function that is based on the loss signal, [Para 0034, Next, the influence of x.sub.ij on model {circumflex over (θ)} is measured by adding a fraction of imp(x.sub.ij|z;{circumflex over (θ)}) scaled by a small value ϵ to the overall loss, and θ.sub.ϵ,x.sub.ij.sub.|z:=argmin.sub.θE.sub.z.sub.i.sub.∈D.sub.train[custom-character(z.sub.i,θ)]+ϵcustom-character(z.sub.−ij;θ)−ϵcustom-character(z,θ) is obtained. The influence of up-weighing the importance of x.sub.ij on {circumflex over (θ)} is obtained as:
PNG
media_image3.png
116
542
media_image3.png
Greyscale
;
Para 0036, Finally, applying the equation immediately above and the chain rule, the influence of x.sub.ij to z′ is obtained as:
PNG
media_image4.png
64
440
media_image4.png
Greyscale
Para 0037, IF.sup.+ measures the influence of a training span on an entire test sequence. Similarly, the influence of a training span to a test span x.sub.kl′ is also measured by applying Equation 3 above to obtain:
PNG
media_image5.png
66
964
media_image5.png
Greyscale
The complete derivation can be found below. IF.sup.+ means IF using a training span on an entire test sequence, and IF.sup.++ means IF using a training span on a test span.]
Zhu is analogous to the claimed invention as they both relate to the implementation of influence function in neural network predictions. 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 Kenji’s teachings to incorporate the teachings of Zhu and provide the one or more computer-readable storage media of claim 1, wherein the operations further comprise: for each entity of the set of entities, determining an entity-embedding based on the machine learning model for each entity of the set of entities, determining an entity-embedding for each training cell of a set of training cells, generating a cell-embedding based on a combination of the entity-embeddings for a subset of the set of entities that is associated with the training cell, determining estimated values for a set of test cells based on the cell- embedding for each of the training cells of the set of training cells; generating a loss signal based on the estimated values for the set of test cells, and for each training cell of the set of training cells, determining the influence metric by employing an influence function that is based on the loss signal [Zhu, Para 0032] to see how the model loss on a test example z′ changes with the training span's importance. Namely, the more important a training span is to z′, the greater this influence score should be, and vice versa.
Regarding claim 16, Kenji further teaches,
A system comprising: a memory device (Para 0232, the HDD 1400); and a processing device (Para 0232, The CPU 1100), operatively coupled to the memory device (Para 0232, operates based on), to perform operations (Para 0232, operates) [Para 0232, The CPU 1100 operates based on a program stored in the ROM 1300 or the HDD 1400, and controls each unit.] comprising:
receiving an incomplete tensor having a multi-dimensional data structure (Fig. 1, DS1) comprising an indexed set of cells (Fig. 1, DT14 and DT33) associated with a set of entities (Para 0014, images) [Para 0014, In FIG. 1, each square in the data set DS1 represents data (image), and the data set DS1 includes a large number of data (images); Para 0016, a dataset DS1 in which each data (image) is associated with a correct answer label indicating the presence or absence of a smile];
training a first tensor-completion model (Para 0016, model M1) using a set of training cells from the incomplete tensor (Para 0016, the data set DS1); [Para 0016, The data adjustment apparatus 100 performs a learning process using the data set DS1 so as to minimize a set loss function, and learns the model M1. The data adjusting apparatus 100 may use various functions as the loss function as long as the influence of each piece of data can be measured in the measurement process];
determining cell-importance metrics (Fig. 1, IV33 and IV14) for the training cells (Para 0016, each piece of data) based on applying an influence function [Fig. 1, MM1, influence function] to estimate an influence of the training cells on minimizing a loss signal (Para 0017, the data adjustment device 100 provides input values (data) to a neural network (model M1), which calculates a predicted value based on the input values, and compares the predicted value with teacher data (correct label) to evaluate the error) determined from training the first tensor-completion model using the training cells (Para 0021) [Para 0016, The data adjusting apparatus 100 may use various functions as the loss function as long as the influence of each piece of data can be measured in the measurement; Para 0020, For example, the data adjustment apparatus 100 may measure the difference between the loss in learning when the entire data set DS1 is used and the loss in learning when data X is removed from data set DS1 as the influence of data X; Para 0021, As shown in the measurement result RS1, the data adjusting device 100 measures the degree of influence that the data DT14 has on the learning of the model M1 as an influence IV14. The influence level IV14 is assumed to be a specific value (for example, 0.2); Para 0025, As shown in the measurement result RS2, the data adjusting device 100 measures the degree of influence that the data DT33 has on the learning of the model M1 as an influence IV33. The influence level IV33 is assumed to be a specific value (for example, 0.7)]
determining, for each entity from the set of entities (Para 0038, pieces of data in the data set DS1), an entity-importance metric (Para 0018, degree of contribution, influence level IV14 and influence level IV33) based on a combination of cell-importance metrics for training cells associated with the entity (Para 0038, device 100 executes similar processing for all pieces of data in the data set DS1) [Para 0018, The degree of influence here indicates that the greater the value, the greater the degree to which the data contributed to the learning of the model M1 (degree of contribution); Para 0021, As shown in the measurement result RS1, the data adjusting device 100 measures the degree of influence that the data DT14 has on the learning of the model M1 as an influence IV14; Para 0025, As shown in the measurement result RS2, the data adjusting device 100 measures the degree of influence that the data DT33 has on the learning of the model M1 as an influence IV33.) that corresponds to the entity; Para 0038, Although FIG. 1 shows processing for only two pieces of data, DT14 and DT33, for the sake of explanation, the data adjustment device 100 executes similar processing for all pieces of data in the data set DS1];
selecting a first cell from the second subset of cells of the incomplete tensor (Para 0028, determines to add data corresponding to the data DT33) based on a sampling (Para 0018, measuring the degree of influence) for a subset of entities (Para 0028, data corresponding to the data DT33 to the data set DS1) associated with the cell [Para 0018, The data adjustment device 100 uses a method for measuring the degree of influence (measurement method MM1) to measure the degree of influence that each data item in the data set DS1 has on the learning of the model M1. The degree of influence here indicates that the greater the value, the greater the degree to which the data contributed to the learning of the model M1 (degree of contribution); Para 0028, Therefore, as indicated by the discrimination result DR2, the data adjustment apparatus 100 determines that the data DT33 has a high contribution to the learning of the model M1, and determines to add data corresponding to the data DT33 to the data set DS1. Then, the data adjustment apparatus 100 adjusts the data set DS1 by adding data corresponding to the data DT33 to the data set DS1. In this way, the data adjustment device 100 updates the data set DS1];
generating an augmented tensor having the multi-dimensional data structure (Para 0039, adjusted dataset DS1) and that includes a union of the incomplete tensor (Para 0039, dataset DS1) with the augmentation tensor [Para 0028, Therefore, as indicated by the discrimination result DR2, the data adjustment apparatus 100 determines that the data DT33 has a high contribution to the learning of the model M1, and determines to add data corresponding to the data DT33 to the data set DS1. Then, the data adjustment apparatus 100 adjusts the data set DS1 by adding data corresponding to the data DT33 to the data set DS1. In this way, the data adjustment device 100 updates the data set DS1; Para 0039, the data adjustment device 100 re-trains the model M1 using an adjusted dataset DS1 in which data with low contribution, such as data DT14, is excluded and data similar to data with high contribution, such as data DT33, is added]; and
adapting one or more parameters of model or second model [Para 0016, The data adjusting apparatus 100 may use various functions as the loss function as long as the influence of each piece of data can be measured in the measurement process; Para 0017, in the backpropagation method, a loss function is used to indicate how far the value of the output layer of the neural network is from the correct state (correct label), and the weights and biases are updated using the steepest descent method or the like so as to minimize the loss function… the data adjustment device 100 provides input values (data) to a neural network (model M1), which calculates a predicted value based on the input values, and compares the predicted value with teacher data (correct label) to evaluate the error].
Kenji does not teach wherein the set of cells comprises a first subset of cells that store a relevant value and a second subset of cells that are empty, store an initialized zero value, or store a null value; generating an augmentation tensor comprising a predicted value for cell based on determining, using model, the predicted value, based at least on one or more other cells of incomplete tensor; a sampling of the set of entities that is weighted proportional to the entity-importance metrics; augmentation tensor comprising cell storing the predicted value; and model being tensor-completion model or tensor array-completion model based at least on the augmented tensor.
Zhu teaches,
a sampling [Para 0008, using training examples from randomly selected batches of the training data D] of the set of entities (Para 0008, training examples, zi) that is weighted proportional to the entity-importance metric (Para 0049, i-specific weighing parameter
PNG
media_image1.png
55
55
media_image1.png
Greyscale
) [Para 0028, calculating importance scores for the training examples; ranking the importance scores (e.g., TracInF(z,z′)) to create a ranked list of the training examples;… For instance, the importance scores TracInF(z,z′) can be calculated as:
PNG
media_image2.png
296
920
media_image2.png
Greyscale
]
Zhu is analogous to the claimed invention as they both relate to the implementation of influence function in neural network predictions. 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 Kenji’s teachings to incorporate the teachings of Zhu and provide set of entities that is weighted by the entity-importance metric for each entity of the set of entities [Zhu, para 0024] to enhance the interpretability of the explanations (i.e., training examples).
Kenji-Zhu teach the above limitations of claim 16 including the first subset of cells, the set of cells (Kenji, Fig. 1 and para 0014), the first cell (Kenji, para 0021), first tensor completion model (Kenji, Para 0021).
Kenji-Zhu do not teach wherein the set of cells comprises a first subset of cells that store a relevant value and a second subset of cells that are empty, store an initialized zero value, or store a null value; generating an augmentation tensor comprising a predicted value for cell based on determining, using model, the predicted value, based at least on one or more other cells of incomplete tensor; augmentation array comprising cell storing the predicted value; and model being tensor-completion model or second tensor-completion model based at least on the augmented tensor.
Tang teaches,
wherein the set of cells comprises a first subset of cells that store a relevant value and a second subset of cells that are empty, store an initialized zero value, or store a null value [Para 0011, Step 2: Utilizing the low-rank property of the received signal tensor model, the received signal tensor model constructed in Step 1 is filled with missing data using a matrix factorization based tensor completion method to obtain the completed tensor model];
generating an augmentation tensor comprising a predicted value for cell based on determining, using model, the predicted value, based at least on one or more other cells of incomplete tensor [Para 0012, Step 3: Expand the completed tensor model to obtain the completed received signal matrix. Use the ESPRIT algorithm for acoustic vector sensor arrays to estimate the source direction of arrival (DOA) of the completed received signal matrix; Para 0064, This invention utilizes a matrix decomposition-based tensor completion algorithm to fold and unfold the received signal data of any acoustic vector sensor array, completing the missing data. This combines tensor completion with traditional DOA estimation algorithms, ultimately achieving accurate DOA estimation even when array element damage results in partial loss of received signal data; Para 0118, The received signal with missing data is filled in using the matrix factorization-based tensor completion algorithm (MFTC) to obtain the complete tensor data. Expanding this tensor yields the complete received signal matrix];
augmentation array comprising cell storing the predicted value [Para 0012, Step 3: Expand the completed tensor model to obtain the completed received signal matrix. Use the ESPRIT algorithm for acoustic vector sensor arrays to estimate the source direction of arrival (DOA) of the completed received signal matrix; Para 0064, This invention utilizes a matrix decomposition-based tensor completion algorithm to fold and unfold the received signal data of any acoustic vector sensor array, completing the missing data. This combines tensor completion with traditional DOA estimation algorithms, ultimately achieving accurate DOA estimation even when array element damage results in partial loss of received signal data; Para 0118, The received signal with missing data is filled in using the matrix factorization-based tensor completion algorithm (MFTC) to obtain the complete tensor data. Expanding this tensor yields the complete received signal matrix]; and
model being tensor-completion model [Para 0064, This invention utilizes a matrix decomposition-based tensor completion algorithm to fold and unfold the received signal data of any acoustic vector sensor array, completing the missing data. This combines tensor completion with traditional DOA estimation algorithms, ultimately achieving accurate DOA estimation even when array element damage results in partial loss of received signal data] or second tensor-completion model based at least on the augmented tensor (alternate).
Regarding claim 20, Kenji-Zhu-Tang teach all the limitations of claim 16 including the first tensor completion model (Kenji, Para 0021).
Kenji further teaches,
wherein the cell-importance metric (Para 0028, IV33) for each training cell (Para 0028, data corresponding to the data DT33 to the data set DS1) is determined by employing an influence function (Fig. 1, MM1/influence function) that is based on the loss signal [Para 0016, The data adjustment apparatus 100 performs a learning process using the data set DS1 so as to minimize a set loss function, and learns the model M1. The data adjusting apparatus 100 may use various functions as the loss function as long as the influence of each piece of data can be measured in the measurement process; Para 0028, In FIG. 1, the data adjusting apparatus 100 determines that the data DT33 is necessary for training the model M1 because the influence level IV33 of the data DT33 is higher than the second threshold TH2.Therefore, as indicated by the discrimination result DR2, the data adjustment apparatus 100 determines that the data DT33 has a high contribution to the learning of the model M1, and determines to add data corresponding to the data DT33 to the data set DS1. Then, the data adjustment apparatus 100 adjusts the data set DS1 by adding data corresponding to the data DT33 to the data set DS1. In this way, the data adjustment device 100 updates the data set DS1.]
Kenji does not teach wherein determining the cell-importance metrics for the training cells comprises: for each entity of the set of entities, determining an entity-embedding based on model; for each training cell of the set of training cells, generating a cell-embedding based on a combination of the entity-embeddings for a subset of the set of entities that is associated with the training cell; determining estimated values for a set of test cells based on the cell- embedding for each of the training cells of the set of training cells; generating the loss signal based on the estimated values for the set of test cells.
Zhu further teaches
for each entity of the set of entities (Para 0033, training span), determining an entity-embedding [Para 0033, According to an exemplary embodiment, the training span is first defined from token i to token j to be xij, and a sequence with xij masked is: x−ij=[x0, . . . ,xi-1,[MASK], . . . ,[MASK],xj+1, . . . ], and its corresponding training data is z−ij];
for each training cell of a set of training cells, generating a cell-embedding (Para 0033, training data is z−ij) based on a combination of the entity-embeddings for a subset of the set of entities that is associated with the training cell [Para 0033, According to an exemplary embodiment, the training span is first defined from token i to token j to be xij, and a sequence with xij masked is: x−ij=[x0, . . . ,xi-1,[MASK], . . . ,[MASK],xj+1, . . . ], and its corresponding training data is z−ij];
determining estimated values [Para 0033, empirical-risk-estimated parameter {circumflex over (θ)} obtained from D.sup.train] for a set of test cells [Para 0037, IF.sup.+ measures the influence of a training span on an entire test sequence…Similarly, the influence of a training span to a test span x.sub.kl′ is also measured] based on the cell- embedding for each of the training cells of the set of training cells; generating (Para 0052, empirical-risk-estimated parameter calculation) a loss signal (Para 0052, loss difference) based on the estimated values for the set of test cells [Para 0052, the logit difference is used as an importance score based on the empirical-risk-estimated parameter {circumflex over (θ)} obtained from D.sup.train as: imp(x.sub.ij|z,{circumflex over (θ)})=logit.sub.y(x;{circumflex over (θ)})−logit.sub.y(x.sub.−ij;{circumflex over (θ)}), wherein every term in the right hand side (RHS) is the logit output evaluated at a model prediction y from model {circumflex over (θ)} to be explained right before applying the SoftMax function, which is a commonly used machine linear function to map input into a probability distribution. This equation indicates how important a training span is, and is equivalent to the loss difference:
imp(z.sub.ij|z,θ)=custom-character(z.sub.−ij;{circumflex over (θ)})−custom-character(z;{circumflex over (θ)}) (3)
when the cross-entropy loss custom-character(z;θ)=−Σ.sub.y.sub.icustom-character(y=y.sub.i)logit.sub.y.sub.i(x;θ) is applied.];
Zhu is analogous to the claimed invention as they both relate to the implementation of influence function in neural network predictions. 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 Kenji’s teachings to incorporate the teachings of Zhu and provide the one or more computer-readable storage media of claim 1, wherein the operations further comprise: for each entity of the set of entities, determining an entity-embedding based on the machine learning model for each entity of the set of entities, determining an entity-embedding for each training cell of a set of training cells, generating a cell-embedding based on a combination of the entity-embeddings for a subset of the set of entities that is associated with the training cell, determining estimated values for a set of test cells based on the cell- embedding for each of the training cells of the set of training cells; generating a loss signal based on the estimated values for the set of test cells [Zhu, Para 0032] to see how the model loss on a test example z′ changes with the training span's importance. Namely, the more important a training span is to z′, the greater this influence score should be, and vice versa.
Regarding claim 22, Kenji-Zhu-Tang teach the limitations of claim 9.
Kenji further teaches,
wherein each entity of the set of entities comprises a portion of the incomplete tensor defined by holding a value of one index of the multi- dimensional data structure constant while varying one or more other indices of the multi- dimensional data structure to vary across an associated range [Para 0021, the data adjusting apparatus 100 may measure the influence of one piece of data (data X) by excluding the data set DS1 and performing re-learning processing. For example, the data adjustment apparatus 100 may measure the difference between the loss in learning when the entire data set DS1 is used and the loss in learning when data X is removed from data set DS1 as the influence of data X].
Claims 2 and 5 is rejected under 35 U.S.C. 103 as being unpatentable over Kenji in view of Zhu and Tang as applied to claim 1 above, and further in view of Zhu et al. (U.S. 20220383096 A1), hereinafter Zhu(2).
Regarding claim 2, Kenji-Zhu-Tang teach all the limitations of claim 1 including the array-completion model (Tang, Para 0064).
Kenji further teaches,
The computer-readable storage medium of claim 1, wherein the operations further comprise: accessing a loss signal (Fig. 1, output from MM1) that was generated during the training (Para 0016, learning process) of the model (Para 0016, M1) [Para 0016, The data adjustment apparatus 100 performs a learning process using the data set DS1 so as to minimize a set loss function, and learns the model M1. The data adjusting apparatus 100 may use various functions as the loss function as long as the influence of each piece of data can be measured in the measurement process]
generating a set of cell-importance cells (Fig. 1, RS1 and RS2), wherein each cell-importance cell stores one of the influence metrics (Fig. 1, IV33 and IV14) that is based on the loss signal (Fig. 1, output of MM1) [Para 0016, The data adjusting apparatus 100 may use various functions as the loss function as long as the influence of each piece of data can be measured in the measurement; Para 0020, For example, the data adjustment apparatus 100 may measure the difference between the loss in learning when the entire data set DS1 is used and the loss in learning when data X is removed from data set DS1 as the influence of data X; Para 0021, As shown in the measurement result RS1, the data adjusting device 100 measures the degree of influence that the data DT14 has on the learning of the model M1 as an influence IV14. The influence level IV14 is assumed to be a specific value (for example, 0.2); Para 0025, As shown in the measurement result RS2, the data adjusting device 100 measures the degree of influence that the data DT33 has on the learning of the model M1 as an influence IV33. The influence level IV33 is assumed to be a specific value (for example, 0.7)); and
for each entity of the set of entities [Para 0038, pieces of data in the data set DS1], determining the entity-importance metric for the entity [Para 0018, degree of contribution, influence level IV14 and influence level IV33] based on a combination of the influence metrics [Para 0038, device 100 executes similar processing for all pieces of data in the data set DS1] that are stored in a subset of the set of cell-importance cells (Fig 1, RS1 and RS2); [Para 0021, As shown in the measurement result RS1, the data adjusting device 100 measures the degree of influence that the data DT14 has on the learning of the model M1 as an influence IV14; Para 0025, As shown in the measurement result RS2, the data adjusting device 100 measures the degree of influence that the data DT33 has on the learning of the model M1 as an influence IV33.) that corresponds to the entity; Para 0038, Although FIG. 1 shows processing for only two pieces of data, DT14 and DT33, for the sake of explanation, the data adjustment device 100 executes similar processing for all pieces of data in the data set DS1]
Kenji-Zhu-Tang does not teach generating a cell-importance array that includes a set of cell-importance cells.
Zhu(2) teaches,
generating a cell-importance array (Para 0051, ∇imp(x'kl|z';
θ
^
) ∇imp(xij|z;
θ
^
))) that includes a set of cell-importance cells (Para 0051, xij) [Para 0051, In step 406, the importance of the training example z on the test example z' is determined using ∇(x'kl|z';
θ
^
)) and ∇imp(xij|z;
θ
^
)). For instance, according to an exemplary embodiment, the importance of the training example z on the test example z' is computed as: ∇imp(x'kl|z';
θ
^
)) ∇imp(xij|z;
θ
^
)).].
Zhu(2) is analogous to the claimed invention as they both relate to the utilization of influence function in neural network predictions. 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 Kenji and Zhu’s teachings to incorporate the teachings of Zhu(2) and provide generating a cell-importance array that includes a set of cell-importance cells to provide commutativity to the set of cell-importance cells via the creation of an array.
Regarding claim 5, Kenji-Zhu-Tang teach all the limitations of claim 1 including the array-completion model (Claim 1: Tang, para 0064).
Kenji does not teach the limitations of claim 5 including for each entity of the set of entities, determining an entity-embedding for each training cell of a set of training cells, generating a cell-embedding based on a combination of the entity-embeddings for a subset of the set of entities that is associated with the training cell, determining estimated values for a set of test cells based on the cell- embedding for each of the training cells of the set of training cells; generating a loss signal based on the estimated values for the set of test cells, and for each training cell of the set of training cells, determining the influence metric by employing an influence function that is based on the loss signal.
Zhu further teaches,
for each entity of the set of entities (Para 0033, training span), determining an entity-embedding [Para 0033, According to an exemplary embodiment, the training span is first defined from token i to token j to be xij, and a sequence with xij masked is: x−ij=[x0, . . . ,xi-1,[MASK], . . . ,[MASK],xj+1, . . . ], and its corresponding training data is z−ij];
for each training cell of a set of training cells, generating a cell-embedding (Para 0033, training data is z−ij) based on a combination of the entity-embeddings for a subset of the set of entities that is associated with the training cell [Para 0033, According to an exemplary embodiment, the training span is first defined from token i to token j to be xij, and a sequence with xij masked is: x−ij=[x0, . . . ,xi-1,[MASK], . . . ,[MASK],xj+1, . . . ], and its corresponding training data is z−ij];
determining estimated values [Para 0033, empirical-risk-estimated parameter {circumflex over (θ)} obtained from D.sup.train] for a set of test cells [Para 0037, IF.sup.+ measures the influence of a training span on an entire test sequence…Similarly, the influence of a training span to a test span x.sub.kl′ is also measured] based on the cell- embedding for each of the training cells of the set of training cells; generating (Para 0052, empirical-risk-estimated parameter calculation) a loss signal (Para 0052, loss difference) based on the estimated values for the set of test cells [Para 0052, the logit difference is used as an importance score based on the empirical-risk-estimated parameter {circumflex over (θ)} obtained from D.sup.train as: imp(x.sub.ij|z,{circumflex over (θ)})=logit.sub.y(x;{circumflex over (θ)})−logit.sub.y(x.sub.−ij;{circumflex over (θ)}), wherein every term in the right hand side (RHS) is the logit output evaluated at a model prediction y from model {circumflex over (θ)} to be explained right before applying the SoftMax function, which is a commonly used machine linear function to map input into a probability distribution. This equation indicates how important a training span is, and is equivalent to the loss difference:
imp(z.sub.ij|z,θ)=custom-character(z.sub.−ij;{circumflex over (θ)})−custom-character(z;{circumflex over (θ)}) (3)
when the cross-entropy loss custom-character(z;θ)=−Σ.sub.y.sub.icustom-character(y=y.sub.i)logit.sub.y.sub.i(x;θ) is applied];
and for each training cell of the set of training cells, determining the influence metric by employing an influence function that is based on the loss signal [Para 0034, Next, the influence of x.sub.ij on model {circumflex over (θ)} is measured by adding a fraction of imp(x.sub.ij|z;{circumflex over (θ)}) scaled by a small value ϵ to the overall loss, and θ.sub.ϵ,x.sub.ij.sub.|z:=argmin.sub.θE.sub.z.sub.i.sub.∈D.sub.train[custom-character(z.sub.i,θ)]+ϵcustom-character(z.sub.−ij;θ)−ϵcustom-character(z,θ) is obtained. The influence of up-weighing the importance of x.sub.ij on {circumflex over (θ)} is obtained as:
PNG
media_image3.png
116
542
media_image3.png
Greyscale
;
Para 0036, Finally, applying the equation immediately above and the chain rule, the influence of x.sub.ij to z′ is obtained as:
PNG
media_image4.png
64
440
media_image4.png
Greyscale
Para 0037, IF.sup.+ measures the influence of a training span on an entire test sequence. Similarly, the influence of a training span to a test span x.sub.kl′ is also measured by applying Equation 3 above to obtain:
PNG
media_image5.png
66
964
media_image5.png
Greyscale
The complete derivation can be found below. IF.sup.+ means IF using a training span on an entire test sequence, and IF.sup.++ means IF using a training span on a test span.
Zhu is analogous to the claimed invention as they both relate to the implementation of influence function in neural network predictions. 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 Kenji’s teachings to incorporate the teachings of Zhu and provide the one or more computer-readable storage media of claim 1, wherein the operations further comprise: for each entity of the set of entities, determining an entity-embedding based on the machine learning model for each entity of the set of entities, determining an entity-embedding for each training cell of a set of training cells, generating a cell-embedding based on a combination of the entity-embeddings for a subset of the set of entities that is associated with the training cell, determining estimated values for a set of test cells based on the cell- embedding for each of the training cells of the set of training cells; generating a loss signal based on the estimated values for the set of test cells, and for each training cell of the set of training cells, determining the influence metric by employing an influence function that is based on the loss signal [Zhu, Para 0032] to see how the model loss on a test example z′ changes with the training span's importance. Namely, the more important a training span is to z′, the greater this influence score should be, and vice versa.
Claim 3 is rejected under 35 U.S.C. 103 as being unpatentable over Kenji in view of Zhu, Tang, and Zhu(2), and further in view of Trayanova et al. (U.S. 20250037874 A1), hereinafter Trayanova.
Regarding claim 3, Kenji-Zhu-Tang-Zhu(2) teach all the limitations of claims 2 including the array-completion model (Claim 1: Tang, para 0064).
Zhu(2) further teaches,
wherein the loss signal includes a set of loss gradients [Para 0048, According to an exemplary embodiment, the importance of the training span xij on training example z=(x,y), i.e., imp(xijlz,θ ̂), is evaluated using a loss gradient].
Zhu(2) is analogous to the claimed invention as they both relate to the utilization of influence functions in neural network predictions. 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 Kenji and Zhu’s teachings to incorporate the teachings of Zhu(2) and provide wherein the loss signal includes a set of loss gradients [Kenji, para 0017] in order to efficiently determine how far an output layer of a neural network is from the correct state.
Kenji-Zhu-Tang-Zhu(2) do not teach a set of gradients are generated during a plurality of epochs during training.
Trayanova teaches,
a set of gradients are generated during a plurality of epochs during training [Para 0070, The maximum number of iterations was 300 for the dense subnetwork and lowered to 100 for the convolutional subnetwork, given its highly increased capacity. Each fold was run using early stopping based on the loss value on a withheld 10% portion of the training fold with a maximum of 2000 epochs (20 gradient updates per epoch).].
Trayanova is analogous to the claimed invention as they both relate to incorporating loss signals to train neural networks. 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 Kenji, Zhu, and Zhu(2)’s teachings to incorporate the teachings of Trayanova and by providing a set of gradients generated during a plurality of epochs during training [Trayanova, para 0070] enables sample parameter configurations from the search space using the Parzen window algorithm to minimize the average validation loss resulting from a stratified ten-times repeated ten-fold cross-validation process.
Claims 12 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Kenji in view of Zhu and Tang, and in further view of Zhu(2) and Trayanova.
Regarding claims 12 and 17, Kenji-Zhu-Tang teaches all the limitations of claims 9 and 16, respectively, including the first tensor-completion model (as in claims 9 and 16, respectively).
Kenji does not teach the limitations of claims 12 and 17 including wherein the loss signal includes a set of loss gradients generated during a plurality of epochs of the training of the first machine learning model.
Zhu(2) further teaches,
wherein the loss signal includes a set of loss gradients [Para 0048, According to an exemplary embodiment, the importance of the training span xij on training example z=(x,y), i.e., imp(xijlz,θ ̂), is evaluated using a loss gradient].
Zhu(2) is analogous to the claimed invention as they both relate to the utilization of influence functions in neural network predictions. 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 Kenji and Zhu’s teachings to incorporate the teachings of Zhu(2) and provide wherein the loss signal includes a set of loss gradients [Kenji, para 0017] in order to efficiently determine how far an output layer of a neural network is from the correct state.
Kenji-Zhu(2) do not teach a set of gradients are generated during a plurality of epochs during training.
Trayanova teaches,
a set of gradients are generated during a plurality of epochs during training [Para 0070, The maximum number of iterations was 300 for the dense subnetwork and lowered to 100 for the convolutional subnetwork, given its highly increased capacity. Each fold was run using early stopping based on the loss value on a withheld 10% portion of the training fold with a maximum of 2000 epochs (20 gradient updates per epoch).].
Trayanova is analogous to the claimed invention as they both relate to incorporating loss signals to train neural networks. 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 Kenji and Zhu(2)’s teachings to incorporate the teachings of Trayanova and by providing a set of gradients generated during a plurality of epochs during training [Trayanova, para 0070] enables sample parameter configurations from the search space using the Parzen window algorithm to minimize the average validation loss resulting from a stratified ten-times repeated ten-fold cross-validation process .
Claim 4 is rejected under 35 U.S.C. 103 as being unpatentable over Kenji in view of Zhu and Tang as applied to claim 1 above, and further in view of Zhu(2) and Pruthi et al. (Estimating Training Data Influence by Tracing Gradient Descent, Published 2020, hereinafter Pruthi).
Regarding claim 4, Kenji-Zhu-Tang teach all the limitations of claim 1, including the array-completion model (Claim 1: Tang, para 0064) and the incomplete array (Kenji, para 0014).
Kenji-Zhu-Tang do not teach the limitations of claim 4, which recites: wherein the operations further comprise training the model by: accessing a set of test cells and a set of training cells from the array; and iteratively: determining estimated values for the set of test cells based on current weights of the model determined using the set of training cells; calculating a loss function based on a comparison between the estimated values for the set of test cells and observed values for the set of test cells; and adjusting the current weights of the machine learning model to decrease loss determined from the loss function.
Zhu(2) further teaches,
wherein the operations further comprise training the model by: accessing a set of test cells (Para 0032, z’) and a set of training cells (Para 0032, z) from the array (Para 0032, D) [Para 0032, classification dataset D={Dtrain, Dtest}… z=(x, y)
∈
Dtrain and z'=(x',y')
∈
Dtest];
Zhu(2) is analogous to the claimed invention as they both relate to the utilization of influence functions in neural network predictions. 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 Kenji and Zhu’s teachings to incorporate the teachings of Zhu(2) and provide the operations further comprise training the machine learning model by: accessing a set of test cells and a set of training cells from the array to determine how far an output layer of a neural network is from the correct state and increase the efficiency and accuracy of a training model [Kenji, para 0017].
Kenji-Zhu-Zhu(2) do not teach iteratively: determining estimated values for the set of test cells based on current weights of the model determined using the set of training cells; calculating a loss function based on a comparison between the estimated values for the set of test cells and observed values for the set of test cells; and adjusting the current weights to decrease loss determined from the loss function
Pruthi teaches,
iteratively: determining estimated values (Sect 3.1, pg. 2, para 2, updating the parameter vector from wt to wt+1) for the set of test cells (Sect 3.1, pg. 2, para 2, Given a set of n training points) based on current weights (Sect 3.1, pg. 2, para 2, by finding parameters w) of the machine learning model (Sect 3.1, pg. 2, para 2, the predictor) determined using the set of training cells (Sect 3.1, pg. 2, para 2, Given a set of n training points) [Sect 3.1, pg. 2, para 2, Given a set of n training points S = {z1, z2,...,zn
∈
Z}, we train the predictor by finding parameters w that minimize the training loss
∑
i
=
1
n
l
w
,
z
i
, via an iterative optimization procedure (such as stochastic gradient descent) which utilizes one training example zt
∈
S in iteration t, updating the parameter vector from wt to wt+1.];
calculating a loss function based on a comparison between the estimated values for the set of test cells (Remark 3.3,
l
w
t
,
z
'
) and observed values (Remark 3.3,
l
w
t
+
1
,
z
'
) for the set of test cells [Remark 3.3, pg. 3, The derivation suggests a way to measure the goodness of the approximation for a given step: We can check that the change in loss for a step
l
w
t
,
z
'
−
l
w
t
+
1
,
z
'
is approximately equal to First-Order Approximation(
B
t
,
z
'
).];
and adjusting the current weights (Sect 3.2, pg. 3, para 1, updating the parameters) to decrease loss determined from the loss function (Sect 4.3, pg. 6, para 4, selecting checkpoints with high loss reduction) [Sect 3.2, pg. 3, para 1, updating the parameters in the training process… we can approximate the change in the loss of a test example in a given iteration t via a simple first-order approximation; Sect 4.3, pg. 6, para 4, We compute the correlation of the influence scores of 100 test points using TracInCP with different checkpoints against the scores from the first-order approximation TracIn. As discussed in Remark D (in the supplementary material), we find that selecting checkpoints with high loss reduction, are more informational than selecting same number of evenly spaced checkpoints.].
Pruthi is analogous to the claimed invention as they both relate to estimating data influence. 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 Kenji, Zhu, and Zhu(2)’s teachings to incorporate the teachings of Pruthi and provide iteratively: determining estimated values for the set of test cells based on current weights of the machine learning model determined using the set of training cells; calculating a loss function based on a comparison between the estimated values for the set of test cells and observed values for the set of test cells; and adjusting the current weights to decrease loss determined from the loss function accessing test cells and training cells and iteratively determining estimated values based on weights, calculating loss function, and adjusting current weights increases the efficiency and accuracy of a machine learning model; Specifically, such method decomposes the difference between the loss of the test point at the end of training versus at the beginning of training along the path taken by the training process [Pruthi, sect 3, para 1].
Claim 6 is rejected under 35 U.S.C. 103 as being unpatentable over Kenji in view of Zhu and Tang as applied to claim 1 above, and further in view of Quan et al. (U.S. 20220261691 A1), hereinafter Quan and He et al. (U.S. 20210312698 A1), hereinafter He.
Regarding claim 6, Kenji-Zhu-Tang teach all the limitations of claim 1 including the first cell (Kenji, para 0021) and the incomplete array (Kenji, para 0014).
Tang further teaches,
wherein cell from the array for which to augment the incomplete array (Para 0012, obtain the completed received signal matrix) with the predicted value is identified (Para 0042, solve only one variable at a time) [Para 0012, Step 3: Expand the completed tensor model to obtain the completed received signal matrix. Use the ESPRIT algorithm for acoustic vector sensor arrays to estimate the source direction of arrival (DOA) of the completed received signal matrix; Para 0042, The alternating direction method is used to solve optimization problems with penalties, that is, to solve only one variable at a time while fixing the other variables as the current optimal values, thus decomposing the optimization problem with penalties into several subproblems; Para 0057-0058, The optimal solution for variable is: Where represents the solution of variable X<sub>(i)</sub> in the (k+1)th iteration, represents the solution of variable X<sub>(i)</sub> in the (k+1)th iteration, and Ω<sup>C</sup> is the set of missing data locations in the tensor].
Tang is analogous to the claimed invention as they both relate to data completion. 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 Kenji’s teachings to incorporate the teachings of Tang and provide augmenting an array with a predicted value in order to [Tang, para 0004] effectively improve estimation algorithms.
Kenji-Zhu-Tang do not teach wherein the operations further comprise: performing the sampling of the set of entities, wherein the sampling is a stochastic sampling of the set of entities; for each dimension of the array, selecting an entity of the set of entities based on the stochastic sampling of the set of entities; and the cell is identified based on a combination of the selected entities for each of the dimensions of the target array.
Quan teaches,
wherein the cell is identified (Para 0046, an original sample z in the original training) based on a combination of the selected entities for each of the dimensions (Para 0046,
PNG
media_image6.png
31
83
media_image6.png
Greyscale
) of the target array [Para 0046, the second training subset may comprise the original samples zl to zn, but it also comprises augmented sample sets that are obtained after the data augmentation processing is performed on one of the original samples. In other words, an original sample z in the original training set may be replaced with the following sample set
PNG
media_image7.png
40
108
media_image7.png
Greyscale
composed of a group of samples obtained after the data augmentation operation is performed on the original sample z].
Quan is analogous to the claimed invention as they both relate to the utilization of influence functions and data augmentation in neural networks. 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 Kenji and Zhu’s teachings to incorporate the teachings of Quan and provide the cell is identified based on a combination of the selected entities for each of the dimensions of the target array as such steps enable data augmentation operation on the original sample for improved training.
Kenji-Zhu-Tang-Quan do not teach wherein the operations further comprise: performing the sampling of the set of entities, wherein the sampling is a stochastic sampling of the set of entities; and for each dimension of the array, selecting an entity of the set of entities based on the stochastic sampling of the set of entities.
He teaches,
wherein the operations further comprise: performing the sampling of the set of entities, wherein the sampling is a stochastic sampling of the set of entities [Para 0080, For instance, the novel-view synthesis system 106 can utilize a stochastic sampling approach (e.g., a priority sampling technique used in reinforcement learning approaches) to sample the subset of image patches {P.sub.i.sup.n}.sub.n=1.sup.N′.]; and
for each dimension of the array (Para 0038, sliding window), selecting an entity (Para 0038, select a subset of portions) of the set of entities (Para 0038, subdivided portions (or regions) of a digital image) based on the stochastic sampling of the set of entities [Para 0038, the term “image patches” refers to subdivided portions (or regions) of a digital image. In particular, the term “image patches” can refer to subdivided portions of a digital image. For instance, the novel-view synthesis system can utilize a sliding window to divide a digital image into a number of portions. Then, the novel-view synthesis system can select a subset of the portions as the image patches (e.g., using stochastic sampling); Para 0081, As shown in FIG. 4, the novel-view synthesis system 106 can sample subsets of image patches from each source image (S.sub.i) belonging to a different viewpoint using stochastic sampling.].
He is analogous to the claimed invention as they both relate to improvements in neural network computations. 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 Kenji, Zhu, and Quan’s teachings to incorporate the teachings of He and provide the operations further comprise: performing the sampling of the set of entities, wherein the sampling is a stochastic sampling of the set of entities for each dimension of the array, selecting an entity of the set of entities based on the stochastic sampling of the set of entities stochastic sampling is a technique in which the image is sampled at appropriate nonuniformly spaced locations rather than at regularly spaced locations, therefore eliminating all forms of aliasing, including unruly forms such as highlight aliasing.
Claims 13 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Kenji in view of Zhu and Tang, and in further view of Zhu(2) and Pruthi.
Regarding claim 13, Kenji-Zhu-Tang teaches all the limitations of claim 9, including the first tensor-completion model (as in claim 9).
Kenji does not teach the remaining limitations of claim 13 such as: the method further comprising training the model by: accessing a set of training cells and a set of test cells from the incomplete tensor; and iteratively: determining estimated values for the set of test cells based on current weights of a model; calculating the loss function based on a comparison between the estimated values for the set of test cells and observed values for the set of test cells; and adjusting the current weights of the model to decrease loss determined from the loss function.
Zhu(2) further teaches,
The method further comprising training the model by: accessing a set of training cells (Para 0032, z) and a set of test cells (Para 0032, z’) from the tensor (Para 0032, D) [Para 0032, classification dataset D={Dtrain, Dtest}… z=(x, y)
∈
Dtrain and z'=(x',y')
∈
Dtest];
Zhu(2) is analogous to the claimed invention as they both relate to the utilization of influence functions in neural network predictions. 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 Kenji and Zhu’s teachings to incorporate the teachings of Zhu(2) and provide the operations further comprise training the machine learning model by: accessing a set of training cells and a set of test cells from the tensor to determine how far an output layer of a neural network is from the correct state and increase the efficiency and accuracy of a training model [Kenji, para 0017].
Kenji-Zhu(2) do not teach iteratively: determining estimated values for the set of test cells based on current weights of model; calculating the loss function based on a comparison between the estimated values for the set of test cells and observed values for the set of test cells; and adjusting the current weights of the model to decrease loss determined from the loss function.
Pruthi further teaches,
iteratively: determining estimated (Sect 3.1, pg. 2, para 2, updating the parameter vector from wt to wt+1) values for the set of test cells (Sect 3.1, pg. 2, para 2, Given a set of n training points) based on current weights (Sect 3.1, pg. 2, para 2, by finding parameters w) of a model (Sect 3.1, pg. 2, para 2, the predictor) [Sect 3.1, pg. 2, para 2, Given a set of n training points S = {z1, z2,...,zn
∈
Z}, we train the predictor by finding parameters w that minimize the training loss
∑
i
=
1
n
l
w
,
z
i
, via an iterative optimization procedure (such as stochastic gradient descent) which utilizes one training example zt
∈
S in iteration t, updating the parameter vector from wt to wt+1.];
calculating the loss function based on a comparison between the estimated values for the set of test cells (Remark 3.3,
l
w
t
,
z
'
) and observed values (Remark 3.3,
l
w
t
+
1
,
z
'
) for the set of test cells [Remark 3.3, pg. 3, The derivation suggests a way to measure the goodness of the approximation for a given step: We can check that the change in loss for a step
l
w
t
,
z
'
−
l
w
t
+
1
,
z
'
is approximately equal to First-Order Approximation(
B
t
,
z
'
).]; and
adjusting the current weights (Sect 3.2, pg. 3, para 1, updating the parameters) of the model to decrease loss determined from the loss function (Sect 4.3, pg. 6, para 4, selecting checkpoints with high loss reduction) [Sect 3.2, pg. 3, para 1, updating the parameters in the training process… we can approximate the change in the loss of a test example in a given iteration t via a simple first-order approximation; Sect 4.3, pg. 6, para 4, We compute the correlation of the influence scores of 100 test points using TracInCP with different checkpoints against the scores from the first-order approximation TracIn. As discussed in Remark D (in the supplementary material), we find that selecting checkpoints with high loss reduction, are more informational than selecting same number of evenly spaced checkpoints.].
Pruthi is analogous to the claimed invention as they both relate to estimating data influence. 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 Kenji, Zhu, and Zhu(2)’s teachings to incorporate the teachings of Pruthi and provide iteratively: determining estimated values for the set of test cells based on current weights of the first machine learning model; calculating a loss function based on a comparison between the estimated values for the set of test cells and observed values for the set of test cells; and adjusting the current weights of the machine learning model to decrease loss determined from the loss function to decrease loss determined from the loss function accessing test cells and training cells and iteratively determining estimated values based on weights, calculating loss function, and adjusting current weights increases the efficiency and accuracy of a machine learning model; Specifically, such method decomposes the difference between the loss of the test point at the end of training versus at the beginning of training along the path taken by the training process [Pruthi, sect 3, para 1].
Claim 19 is a system claim that recites the same limitations as claim 13. Therefore, claim 19 is rejected using the same rationale as claim 13.
Claims 15 and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Kenji in view of Zhu and Tang, and in further view of He and Quan.
Regarding claim 15, Kenji-Zhu-Tang teaches all the limitations of claim 9 including the second subset of cells (Kenji, Para 0021, DT14) and the incomplete tensor (Kenji, Para 0016).
Kenji does not teach performing stochastic sampling of the set of entities; for each dimension of the tensor, selecting an entity of the set of entities based on the stochastic sampling of the set of entities; and wherein the cell from the second subset of cells is selected based on a combination of the selected entities for each of the dimensions of the target array.
He further teaches,
performing stochastic sampling of the set of entities [Para 0080, For instance, the novel-view synthesis system 106 can utilize a stochastic sampling approach (e.g., a priority sampling technique used in reinforcement learning approaches) to sample the subset of image patches {P.sub.i.sup.n}.sub.n=1.sup.N′.];
for each dimension of the tensor (Para 0038, sliding window), selecting an entity(Para 0038, select a subset of portions) of the set of entities (Para 0038, subdivided portions (or regions) of a digital image) based on the stochastic sampling of the set of entities [Para 0038, the term “image patches” refers to subdivided portions (or regions) of a digital image. In particular, the term “image patches” can refer to subdivided portions of a digital image. For instance, the novel-view synthesis system can utilize a sliding window to divide a digital image into a number of portions. Then, the novel-view synthesis system can select a subset of the portions as the image patches (e.g., using stochastic sampling); Para 0081, As shown in FIG. 4, the novel-view synthesis system 106 can sample subsets of image patches from each source image (S.sub.i) belonging to a different viewpoint using stochastic sampling.].
He is analogous to the claimed invention as they both relate to improvements in neural network computations. 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 Kenji’s teachings to incorporate the teachings of He and provide performing stochastic sampling of the set of entities; for each dimension of the tensor, selecting an entity of the set of entities based on the stochastic sampling of the set of entities; stochastic sampling is a technique in which the image is sampled at appropriate nonuniformly spaced locations rather than at regularly spaced locations, therefore eliminating all forms of aliasing, including unruly forms such as highlight aliasing.
He does not teach wherein the cell from the second subset of cells is selected based on a combination of the selected entities for each of the dimensions of a target array.
Quan further teaches,
wherein the cell from subset of cells is selected (Para 0046, an original sample z in the original training) based on a combination of the selected entities (Para 0046,
PNG
media_image6.png
31
83
media_image6.png
Greyscale
) for each of the dimensions of a target array.
[Para 0046, the second training subset may comprise the original samples zl to zn, but it also comprises augmented sample sets that are obtained after the data augmentation processing is performed on one of the original samples. In other words, an original sample z in the original training set may be replaced with the following sample set
PNG
media_image7.png
40
108
media_image7.png
Greyscale
composed of a group of samples obtained after the data augmentation operation is performed on the original sample z.].
Quan is analogous to the claimed invention as they both relate to the utilization of influence functions and data augmentation in neural networks. 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 Kenji and He’s teachings to incorporate the teachings of Quan and provide wherein the cell from subset of cells is selected based on a combination of the selected entities for each of the dimensions of the target array as such steps enable data augmentation operation on the original sample for improved training.
Claim 18 is a system claim that recites the same limitations of claim 15. Therefore, claim 18 is rejected using the same rationale as claim 15.
Claim 21 is rejected under 35 U.S.C. 103 as being unpatentable over Kenji in view of Zhu and Tang, and in further view of Wang et al. (US 12141037 B2), hereinafter Wang
Regarding claim 21, Kenji-Zhu-Tang teach the limitations of claim 1.
Wang teaches,
wherein an individual cell of the multi-dimensional data structure comprises a plurality of embeddings, wherein each embedding is associated with a dimension of the incomplete array [Col 3, lines 28-36, The input data includes the values (X.sub.Ω) and positions/indexes (Ω) of the observed entries (first entries) of the incomplete matrix identified from the object data OD, a first parameter (I.sub.m), a second parameter (ζ) and a third parameter (δ.sub.TH). The first parameter (I.sub.m) is a maximum total number of one or more outer optimization iterations performed by the analysis model, the second parameter (ζ) is a tolerance coefficient, and the third parameter (δ.sub.TH) is a convergence coefficient].
Wang is analogous to the claimed invention as they both relate to matrix completion. 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 Kenji and Zhu’s teachings to incorporate the teachings of Wang and provide a plurality of embeddings in order to improve model performance via dimensionality reduction.
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