Data analysis device, information processing system, and data analysis method

Abstract

The data analysis device of the present application is provided with a task processing section (121) having a learning processing section (1211) that inputs a data set provided with a plurality of learning tasks each having a plurality of data, and outputs data representing an inference model parameter or metadata, and a task search section (122) that inputs the data set and data representing the output result of the task processing section (121), and outputs data representing the contribution degree of the learning task possessed by the data set with respect to the output result, or data representing the learning task having a characteristic in the contribution degree.

Description

Technical Field

[0001] This disclosure relates to data analysis apparatus, information processing system, and data analysis method for analyzing data. Background Technology

[0002] Patent Document 1 discloses a method for analyzing factors that influence predictions based on a learned machine learning model and reconstructing the learning dataset to re-implement the learning process. Specifically, Patent Document 1 demonstrates a sensitivity analysis of the impact of changes in the learning data on predictions performed by a retrieval unit, a confirmation unit indicating data with a high degree of influence and soliciting the user's judgment, and a configuration unit reconstructing the data to create data for relearning.

[0003] Patent Document 1: Japanese Patent Application Publication No. 2022-131406

[0004] On the other hand, when it is desirable to improve the reliability of devices that utilize multi-task learning or meta-learning, it is expected that the reliability evaluation will be carried out on a task-by-task basis.

[0005] However, the prior art disclosed in Patent Document 1 is not intended to improve reliability. Furthermore, even if it were used for that purpose, it cannot be evaluated on a task-by-task basis, but must be judged by observing the contribution from individual data. Summary of the Invention

[0006] This disclosure was made to solve the aforementioned problems, and its purpose is to provide a data analysis device capable of analyzing data on a task-by-task basis.

[0007] The data analysis apparatus disclosed herein is characterized by comprising: a task processing unit having a learning processing unit that takes into input a dataset having multiple learning tasks, each having multiple data points, and outputs data representing inference model parameters or metadata; and a task retrieval unit that takes into input a dataset and data representing the output result of the task processing unit, and outputs data representing the contribution of the learning tasks in the dataset to the output result, or data representing the learning tasks with characteristics of the contribution.

[0008] According to this disclosure, it is configured as described above, thus enabling data analysis on a task-by-task basis. Attached Figure Description

[0009] Figure 1 This is a block diagram illustrating a structural example of an information processing system equipped with the data analysis apparatus of Embodiment 1.

[0010] Figure 2 This is a block diagram illustrating a structural example of the data analysis apparatus in Embodiment 1.

[0011] Figure 3This is a block diagram illustrating another structural example of the data analysis apparatus in Embodiment 1.

[0012] Figure 4 This is a block diagram illustrating another structural example of the data analysis apparatus in Embodiment 1.

[0013] Figure 5 This is a block diagram illustrating another structural example of the data analysis apparatus in Embodiment 1.

[0014] Figure 6 This is a flowchart illustrating an example of the operation of the data analysis device in Implementation Method 1.

[0015] Figure 7 This is a diagram illustrating an example of the dataset and reasoning task (testing task) used in the data analysis apparatus of Embodiment 1.

[0016] Figure 8 This is a diagram illustrating an example of operation in the case of multi-task learning of the task processing unit in Embodiment 1.

[0017] Figure 9 This is a diagram illustrating an example of operation in the case of meta-learning of the task processing unit in Implementation 1.

[0018] Figure 10 This is a diagram illustrating an example of the contribution of the computational learning task to the learning processing result in the task retrieval unit of Implementation 1.

[0019] Figure 11 This is a diagram illustrating an example of the contribution of the computational learning task to the reasoning processing result in the task retrieval unit of Implementation 1.

[0020] Figure 12 This is a diagram illustrating an example of the contribution of the computational learning task of the task retrieval unit in Implementation 1 to the additional adjustment processing result.

[0021] Figure 13 This is a diagram illustrating an example of the contribution of a computational learning task to the reasoning processing result based on additional adjustment processing of the task retrieval unit in Implementation 1.

[0022] Figure 14A , Figure 14B This is a diagram illustrating an example of the operation of the data analysis device in Embodiment 1. Figure 14A This is a diagram illustrating the current state of the technology. Figure 14B This is a diagram illustrating the data analysis apparatus of Embodiment 1.

[0023] Figure 15 This is a block diagram illustrating a structural example of the data analysis device in Embodiment 2.

[0024] Figure 16 This is a block diagram illustrating another structural example of the data analysis device in Embodiment 2.

[0025] Figure 17 This is a block diagram illustrating a structural example of the data analysis device in Embodiment 3.

[0026] Figure 18A , Figure 18B This is a block diagram illustrating an example of the hardware structure of the data analysis device in embodiments 1 to 3. Detailed Implementation

[0027] Hereinafter, the embodiments will be described in detail with reference to the accompanying drawings.

[0028] Implementation method 1.

[0029] Figure 1 This is a block diagram illustrating a structural example of an information processing system 1 equipped with the data analysis device 12 of Embodiment 1.

[0030] For example, such as Figure 1 As shown, the information processing system 1 includes a dataset acquisition unit 11, a data analysis device 12, and an information processing unit 13. For example, a text image classification system that performs text image classification can be cited as an example of this information processing system 1.

[0031] The dataset acquisition unit 11 acquires the dataset. For example, such as... Figure 2 As shown, the dataset contains multiple learning tasks. Furthermore, each learning task has multiple datasets.

[0032] The data analysis device 12 takes into account the dataset acquired by the dataset acquisition unit 11 and analyzes the data in the dataset on a task-by-task basis. An example of the structure of this data analysis device 12 will be described later.

[0033] The information processing unit 13 performs information processing on a task-by-task basis based on the analysis results of the data analysis device 12. For example, if the information processing system 1 is a text-image classification system, the information processing unit 13 classifies the text images represented by the data input to the information processing system 1 on a task-by-task basis.

[0034] Next, refer to Figure 2 An example of the structure of the data analysis device 12 in Embodiment 1 will be described.

[0035] For example, such as Figure 2 As shown, the data analysis device 12 includes a task processing unit 121 and a task retrieval unit 122.

[0036] For example, such as Figure 2As shown, the task processing unit 121 includes a learning processing unit 1211 and a reasoning processing unit 1212. Figure 2 Multi-task learning is performed in the task processing unit 121 shown.

[0037] The learning processing unit 1211 takes the dataset as input, performs learning processing based on the dataset, and outputs data representing the parameters of the inference model. Figure 2 In the example, the learning processing unit 1211 receives a dataset containing learning tasks 1 to M. Additionally, in... Figure 2 In the example, learning task 1 has data 1-1 to 1-3, and learning task M has data M-1 to M-3.

[0038] The inference processing unit 1212 takes into input data representing the output result of the learning processing unit 1211 and data representing the inference task, performs inference processing based on the data representing the output result and the inference data, and outputs data representing the inference processing result relative to the inference task or data representing the performance evaluation value of the inference processing result. The inference task is a test task or a purpose task.

[0039] In addition, Figure 2 The image shows the learning and reasoning processes performed by the task processing unit 121.

[0040] However, it is not limited to this. For example, the task processing unit 121 may also perform reasoning processing after learning processing and additional adjustment processing. That is, in this case, for example, as... Figure 3 As shown, in addition to the learning processing unit 1211 and the reasoning processing unit 1212, the task processing unit 121 also has an adjustment processing unit 1213. Figure 3 Multi-task learning or meta-learning is performed in the task processing unit 121 shown.

[0041] In this case, the learning processing unit 1211 takes a dataset as input, performs learning processing based on the dataset, and outputs data representing inference model parameters or data representing meta-parameters. That is, the learning processing unit 1211 outputs data representing inference model parameters when performing multi-task learning, and outputs data representing meta-parameters when performing meta-learning.

[0042] The adjustment processing unit 1213 takes into input data representing the output result of the learning processing unit 1211 and data representing the inference task. Based on the data representing the output result and the inference task, it performs additional adjustment processing and outputs data representing the inference model parameters. Here, if the output result of the learning processing unit 1211 is data representing the inference model parameters, the adjustment processing unit 1213 adjusts the inference model parameters in the additional adjustment processing and outputs data representing the adjusted inference model parameters. Alternatively, if the output result of the learning processing unit 1211 is a meta-parameter, the adjustment processing unit 1213 determines the inference model parameters in the additional adjustment processing and outputs data representing the inference model parameters.

[0043] In addition, the inference processing unit 1212 inputs data representing the output result of the adjustment processing unit 1213 and data representing the inference task, performs inference processing based on the data representing the output result and the inference task, and outputs data representing the inference processing result relative to the inference task or data representing the performance evaluation value of the inference processing result.

[0044] Alternatively, for example, the task processing unit 121 may perform only learning processing without reasoning processing. That is, in this case, for example, as... Figure 4 As shown, the task processing unit 121 includes a learning processing unit 1211. In Figure 4 Multi-task learning or meta-learning is performed in the task processing unit 121 shown.

[0045] In this case, the learning processing unit 1211 takes a dataset as input, performs learning processing based on the dataset, and outputs data representing inference model parameters or data representing meta-parameters. That is, the learning processing unit 1211 outputs data representing inference model parameters when performing multi-task learning, and outputs data representing meta-parameters when performing meta-learning.

[0046] Alternatively, for example, the task processing unit 121 may perform only learning processing and additional adjustment processing without performing reasoning processing. That is, in this case, for example, as... Figure 5 As shown, the task processing unit 121 includes a learning processing unit 1211 and an adjustment processing unit 1213. Figure 5 Multi-task learning or meta-learning is performed in the task processing unit 121 shown.

[0047] In this case, the learning processing unit 1211 takes a dataset as input, performs learning processing based on the dataset, and outputs data representing inference model parameters or data representing meta-parameters. That is, the learning processing unit 1211 outputs data representing inference model parameters when performing multi-task learning, and outputs data representing meta-parameters when performing meta-learning.

[0048] The task retrieval unit 122 calculates the contribution of the learning task in the dataset to the output result based on the dataset and the output result of the task processing unit 121.

[0049] Alternatively, the task retrieval unit 122 performs the aforementioned calculation process and a selection process based on the calculation results to select learning tasks with contribution characteristics. In this case, during the selection process, the task retrieval unit 122 selects at least one of the learning tasks that contribute significantly to the output results or those that contribute little to the output results.

[0050] Data representing the contribution calculated by the task retrieval unit 122, or data representing the learning task (identification information) selected by the task retrieval unit 122, is output to the outside. Thus, the user can grasp the aforementioned contribution or learning tasks with contribution characteristics.

[0051] Additionally, "contribution" refers, for example, to the amount of change in the calculation result, or an approximation of such change, when the data or processing method used in the calculation of performance evaluation values ​​or predicted values ​​is altered based on certain rules. Furthermore, if the data is parameterized, the amount of change can also be expressed as a rate of change.

[0052] For example, the aforementioned change could also refer to the change in performance metrics when specific data is removed from the dataset for learning purposes.

[0053] Alternatively, for example, the aforementioned change could also be the change in performance metrics when specific data in the dataset is replaced with dummy data for learning.

[0054] Alternatively, for example, the aforementioned variation could also be the differential coefficient of the performance metric with respect to the weight parameters, in the case of learning after weighting and averaging the data values ​​of specific data and virtual data in the dataset.

[0055] Furthermore, "the contribution of each task" refers, for example, to the amount of change in the computational processing results, or an approximation of such a amount, when a coordinated change is applied to the overall data or processing method of each task. Additionally, if the data is parameterized, the amount of change can be expressed as a rate of change.

[0056] For example, the aforementioned change could also be the change in performance metrics when learning is performed by removing a specific task from the dataset.

[0057] Alternatively, for example, the aforementioned change could also be the change in performance metrics when a specific task in the dataset is replaced with a virtual task for learning.

[0058] Alternatively, for example, the aforementioned variation could also be the differential coefficient of the performance metric with respect to the weight parameters, in the case where all the data contained in a specific task in the dataset are learned by weighting the data values ​​of the virtual data with a common weight parameter.

[0059] In addition, when the task processing unit 121 performs multiple processes, for example, the task retrieval unit 122 calculates the contribution between the input and output of each of the above processes and synthesizes the contribution to calculate the contribution of the learning task to the final output result in the task processing unit 121.

[0060] For example, when the task processing unit 121 performs learning processing and reasoning processing, consider the case where the task retrieval unit 122 calculates the contribution of the learning task to the output of the reasoning processing unit 1212.

[0061] In this case, firstly, the task retrieval unit 122 calculates the contribution of the learning task to the output of the learning processing unit 1211. This contribution is then taken as the first contribution.

[0062] Furthermore, the task retrieval unit 122 calculates the contribution of the output of the learning processing unit 1211 to the output of the reasoning processing unit 1212. This contribution is then used as the second contribution.

[0063] Then, the task retrieval unit 122 synthesizes the first contribution and the second contribution, thereby calculating the contribution of the learning task to the output of the reasoning processing unit 1212.

[0064] In addition, for example, when the task processing unit 121 performs learning processing, additional adjustment processing and reasoning processing, the task retrieval unit 122 may calculate the contribution of the learning task to the output of the reasoning processing unit 1212.

[0065] In this case, firstly, the task retrieval unit 122 calculates the contribution of the learning task to the output of the learning processing unit 1211. This contribution is then taken as the first contribution.

[0066] In addition, the task retrieval unit 122 calculates the contribution of the output of the learning processing unit 1211 to the output of the adjustment processing unit 1213. This contribution is taken as the third contribution.

[0067] In addition, the task retrieval unit 122 calculates the contribution of the output result of the adjustment processing unit 1213 to the output result of the reasoning processing unit 1212. This contribution is used as the fourth contribution.

[0068] Then, the task retrieval unit 122 synthesizes the first contribution, the third contribution, and the fourth contribution, thereby calculating the contribution of the learning task to the output of the reasoning processing unit 1212.

[0069] In addition, for example, when the task processing unit 121 performs learning processing and adjustment processing, the task retrieval unit 122 may calculate the contribution of the learning task to the output result of the adjustment processing unit 1213.

[0070] In this case, firstly, the task retrieval unit 122 calculates the contribution of the learning task to the output of the learning processing unit 1211. This contribution is then taken as the first contribution.

[0071] In addition, the task retrieval unit 122 calculates the contribution of the output of the learning processing unit 1211 to the output of the adjustment processing unit 1213. This contribution is taken as the third contribution.

[0072] Then, the task retrieval unit 122 synthesizes the first contribution and the third contribution, thereby calculating the contribution of the learning task to the output of the adjustment processing unit 1213.

[0073] Next, refer to Figure 6 An example of the operation of the data analysis device 12 in Embodiment 1 will be described.

[0074] In addition, the following describes the task processing unit 121 as follows: Figure 2 The operation example under the structure shown will be explained, but the task processing unit 121 is... Figures 3-5 The operation example in the case of the structure shown is the same. In addition, the following shows the case where the task retrieval unit 122 selects a learning task with contribution characteristics.

[0075] In an example of the operation of the data analysis device 12 in Embodiment 1, for example, such as Figure 6 As shown, the task processing unit 121 performs learning processing and inference processing (step ST101). Specifically, the learning processing unit 1211 receives a dataset containing multiple learning tasks, each with multiple data points, performs learning processing based on this dataset, and outputs data representing the parameters of the inference model. Then, the inference processing unit 1212 receives data representing the output result of the learning processing unit 1211 and data representing the inference task, performs inference processing based on the data representing the output result and the inference task, and outputs data representing the inference processing result relative to the inference task or a performance evaluation value representing the inference processing result.

[0076] Next, the task retrieval unit 122 performs calculation processing based on the dataset and the output of the task processing unit 121 to calculate the contribution of the learning task of the dataset to the output (step ST102).

[0077] At this point, the task retrieval unit 122 first calculates the contribution of the learning task to the output of the learning processing unit 1211. This contribution is then taken as the first contribution.

[0078] Furthermore, the task retrieval unit 122 calculates the contribution of the output of the learning processing unit 1211 to the output of the reasoning processing unit 1212. This contribution is then used as the second contribution.

[0079] Then, the task retrieval unit 122 synthesizes the first contribution and the second contribution, thereby calculating the contribution of the learning task to the output of the reasoning processing unit 1212.

[0080] Next, the task retrieval unit 122 performs a selection process based on the above calculation results to select learning tasks with characteristics of contribution (step ST103). At this time, in the selection process, the task retrieval unit 122 selects at least one of the learning tasks that contribute greatly to the above output results or the learning tasks that contribute little to the above output results.

[0081] The data indicating the learning task (identification information) selected by the task retrieval unit 122 is output externally. This allows the user to understand the aforementioned learning tasks with contribution levels or contribution characteristics.

[0082] Figure 7 This is a diagram illustrating an example of the dataset and reasoning task (testing task) used in the data analysis apparatus 12 of Embodiment 1. Figure 7 The example shown is a dataset used to classify various text images.

[0083] In addition, Figure 7 In the case where the dataset is used for multi-task learning, the learning tasks in the dataset only have... Figure 7 The learning data is on the left. Additionally, in Figure 7 In the case where the dataset is a meta-learning dataset, the learning task of the dataset has Figure 7 The left side represents the learning data, and the right side represents the test data. That is, in the case of meta-learning, it is necessary to learn meta-parameters to improve the test performance of the learning results for each task. Therefore, the learning task includes test data in addition to the learning data.

[0084] In addition, each type of image with text is categorized in each learning task.

[0085] In addition, such as Figure 7 As shown, the types of tasks typically used for learning differ from those used for the purpose of learning or for testing.

[0086] Figure 8 This diagram illustrates an example of operation of the task processing unit 121 in Embodiment 1 during multi-task learning. Furthermore, the additional adjustment processing is often referred to as "additional learning" or "fine-tuning."

[0087] Figure 8 The processing flow shown is from left to right.

[0088] exist Figure 8 In the first stage, it is shown, for example, as follows: Figure 4 The task processing unit 121 shown only performs learning processing.

[0089] In this case, the learning processing unit 1211 inputs a dataset, determines the parameters of the inference model based on the learning data in the dataset, and outputs data representing the parameters of the inference model.

[0090] The second stage is illustrated as follows: Figure 2 The task processing unit 121 shown is performing learning processing and reasoning processing.

[0091] In this case, the learning processing unit 1211 inputs a dataset, determines the parameters of the inference model based on the learning data in the dataset, and outputs data representing the parameters of the inference model.

[0092] In addition, the inference processing unit 1212 takes in data representing inference model parameters and data representing inference tasks output by the learning processing unit 1211, performs performance evaluation of inference processing or inference processing results based on the inference model parameters and the inference tasks, and outputs data representing the performance evaluation value of inference processing or inference processing results.

[0093] The third stage is illustrated for example, as shown in... Figure 5 The task processing unit 121 shown is performing learning processing and additional adjustment processing.

[0094] In this case, the learning processing unit 1211 inputs a dataset, determines the parameters of the inference model based on the learning data in the dataset, and outputs data representing the parameters of the inference model.

[0095] In addition, the adjustment processing unit 1213 inputs data representing the parameters of the reasoning model and data representing the reasoning task output by the learning processing unit 1211, improves the parameters of the reasoning model based on the reasoning task, and outputs data representing the parameters of the reasoning model.

[0096] The fourth stage shows, for example, Figure 3 The task processing unit 121 shown performs learning processing, additional adjustment processing, and reasoning processing.

[0097] In this case, the learning processing unit 1211 inputs a dataset, determines the parameters of the inference model based on the learning data in the dataset, and outputs data representing the parameters of the inference model.

[0098] In addition, the adjustment processing unit 1213 inputs data representing the parameters of the reasoning model and data representing the reasoning task output by the learning processing unit 1211, improves the parameters of the reasoning model, and outputs data representing the parameters of the reasoning model.

[0099] In addition, the inference processing unit 1212 takes in data representing inference model parameters and data representing inference tasks output by the adjustment processing unit 1213, performs performance evaluation of inference processing or inference processing results based on the inference model parameters and the inference tasks, and outputs data representing the performance evaluation value of inference processing or inference processing results.

[0100] Figure 9 This is a diagram illustrating an example of operation in the case of meta-learning in the task processing unit 121 of Embodiment 1. Furthermore, learning processing is generally referred to as "meta-learning," and additional adjustment processing is generally referred to as "adaptation" or "learning."

[0101] Figure 9 The processing flow shown is from left to right.

[0102] Furthermore, in the case of meta-learning, the additional adjustment processing in the previous stage is necessary when performing inference processing, therefore it is not related to... Figure 8 The second phase, as shown, is a significant part of the multi-task learning scenario.

[0103] exist Figure 8 In the first stage, it is shown, for example, as follows: Figure 4 The task processing unit 121 shown only performs learning processing.

[0104] In this case, the learning processing unit 1211 inputs a dataset, determines the parameters of the inference model based on the learning data in the dataset, determines the meta-parameter that has the best inference performance for the test data, and outputs data representing the meta-parameter.

[0105] The third stage is illustrated for example, as shown in... Figure 5 The task processing unit 121 shown is performing learning processing and additional adjustment processing.

[0106] In this case, the learning processing unit 1211 inputs a dataset, determines the parameters of the inference model based on the learning data in the dataset, determines the meta-parameters that have the best inference performance based on the test data, and outputs data representing the meta-parameters.

[0107] In addition, the adjustment processing unit 1213 inputs the data representing the meta-parameters output by the learning processing unit 1211 and the data representing the reasoning task, determines the inference model parameters based on the meta-parameters and the inference task, and outputs the data representing the inference model parameters.

[0108] The fourth stage shows, for example, Figure 3 The task processing unit 121 shown performs learning processing, additional adjustment processing, and reasoning processing.

[0109] In this case, the learning processing unit 1211 inputs a dataset, determines the parameters of the inference model based on the learning data in the dataset, determines the meta-parameters that have the best inference performance based on the test data, and outputs data representing the meta-parameters.

[0110] In addition, the adjustment processing unit 1213 inputs the data representing the meta-parameters output by the learning processing unit 1211 and the data representing the reasoning task, determines the inference model parameters based on the meta-parameters and the inference task, and outputs the data representing the inference model parameters.

[0111] In addition, the inference processing unit 1212 takes in data representing inference model parameters and data representing inference tasks output by the adjustment processing unit 1213, performs performance evaluation of inference processing or inference processing results based on the inference model parameters and the inference tasks, and outputs data representing the performance evaluation value of inference processing or inference processing results.

[0112] Figure 10 This is a diagram illustrating an example of the calculation operation of the contribution of the learning task in the task retrieval unit 122 of Embodiment 1 to the learning processing result.

[0113] like Figure 10 As shown, for example, when calculating the contribution of the learning task to the learning processing result, the task retrieval unit 122 calculates the differential coefficients of the perturbation parameters, i.e., the sensitivity matrix A. In this case, the size of the contribution calculated by the task retrieval unit 122, i.e., the sensitivity matrix A, is <the number of inference model parameters> × <the number of learning tasks>.

[0114] That is, for example, the task processing unit 121 determines the parameters of the inference model based on the loss function of the entire data set. The loss function of the entire data set is obtained by synthesizing the loss function of each learning task in the dataset according to each learning task.

[0115] Furthermore, in this case, the task retrieval unit 122 calculates the differential coefficient with respect to the perturbation parameter of the loss function of the entire dataset, based on the loss function of any learning task transformed by the perturbation parameter common to the data in the learning task, and uses the differential coefficient as the contribution of the learning task.

[0116] Here, we consider using inference model parameters (θ) and a dataset (D1, ..., D2) with M learning tasks. M The perturbation of the loss function (C) with respect to the data as a whole (x) is expressed as follows (1), and the loss function (L) for each learning task.

[0117]

[0118] Furthermore, for equation (1), consider the loss function (C') when any learning task (j) in the dataset is given a perturbation caused by a perturbation parameter (ε) as shown in equation (2).

[0119]

[0120] Here, the value of θ that minimizes C' depends on ε. Therefore, the task retrieval unit 122 uses the value of this differential coefficient at ε=0 as the contribution of the learning task (j) to the learning processing result. Then, the task retrieval unit 122 calculates the contribution of each learning task by performing the above processing on all learning tasks in the dataset.

[0121] Furthermore, the output of the learning process can be expressed as an implicit function of the perturbation parameters, for example, through minimizing the loss function. Therefore, the differential coefficients between the variables that provide the solution to the minimization problem can be computed using implicit function differentiation, as is the case with existing techniques.

[0122] That is, for example, the task processing unit 121 determines the value of the extreme value of the loss function that realizes the entire data set, or the model parameters that serve as an approximation of that value.

[0123] Furthermore, in this case, the task retrieval unit 122 calculates the contribution of the learning task to the inference model parameters determined by the task processing unit 121 through implicit function differentiation.

[0124] Here, in the prior art, as shown in equation (3) below, the perturbation of the loss function is considered on a single data point (x) instead of distinguishing data according to the learning task.

[0125] In contrast, in the data analysis apparatus 12 of Embodiment 1, as shown in Equation (2), the perturbation of the loss function is considered for each learning task (D). That is, in the data analysis apparatus 12 of Embodiment 1, multiple data are bundled into one learning task, and the overall coordination perturbation of the data is considered.

[0126]

[0127] In addition, Figure 10 The diagram shows an example of the task retrieval unit 122 calculating the differential coefficients using numerical differentiation.

[0128] However, it is not limited to this; the task retrieval unit 122 can also use various mathematical facts to calculate precise differential coefficients.

[0129] In addition, when the task processing unit 121 performs multiple processes, for example, the task retrieval unit 122 calculates differential coefficients as the contribution between the input and output in the process for each process performed by the task processing unit 121 and forms a matrix, and synthesizes the contribution by calculating the matrix product of the matrices of the multiple processes.

[0130] For example, such as Figure 11 As shown, when the task processing unit 121 performs learning processing and inference processing, the task retrieval unit 122 calculates a sensitivity matrix B, which is a differential coefficient, in addition to the sensitivity matrix A, as the contribution of the output result of the learning processing unit 1211 to the output result of the inference processing unit 1212.

[0131] At this point, the contribution calculated by the task retrieval unit 122, i.e., the size of the sensitivity matrix B, is <the number of performance evaluation values> × <the number of inference model parameters>.

[0132] And, for example, such as Figure 11 As shown, the task retrieval unit 122 calculates the matrix product of sensitivity matrix A and sensitivity matrix B. The contribution of each learning task to the output of the inference processing unit 1212 is calculated. The magnitude of this contribution is a matrix of <number of performance evaluation values> × <number of learning tasks>.

[0133] Additionally, for example, such as Figure 12As shown, when the task processing unit 121 performs learning processing and additional adjustment processing, the task retrieval unit 122 calculates the sensitivity matrix C, which is a differential coefficient, in addition to the sensitivity matrix A, as the contribution of the output result of the learning processing unit 1211 to the output result of the adjustment processing unit 1213.

[0134] At this time, the contribution degree calculated by the task retrieval unit 122, i.e. the size of the sensitivity matrix C, is <number of inference model parameters> × <number of inference model parameters>.

[0135] And, for example, such as Figure 12 As shown, the task retrieval unit 122 calculates the matrix product of the sensitivity matrix A and the sensitivity matrix C. The contribution of each learning task to the output of the adjustment processing unit 1213 is calculated. The magnitude of this contribution is a matrix of <number of inference model parameters> × <number of learning tasks>.

[0136] Additionally, for example, such as Figure 13 As shown, when the task processing unit 121 performs learning processing, additional adjustment processing, and inference processing, the task retrieval unit 122 calculates a sensitivity matrix B' as a differential coefficient, in addition to the sensitivity matrix A and sensitivity matrix C, as the contribution of the output result of the adjustment processing unit 1213 to the output result of the inference processing unit 1212.

[0137] At this time, the contribution calculated by the task retrieval unit 122, i.e. the size of the sensitivity matrix B', is <the number of performance evaluation values> × <the number of inference model parameters>.

[0138] And, for example, such as Figure 13 As shown, the task retrieval unit 122 calculates the matrix product of sensitivity matrix A, sensitivity matrix C, and sensitivity matrix B'. The contribution of each learning task to the output of the inference processing unit 1212 is calculated. The magnitude of this contribution is a matrix of <number of performance evaluation values> × <number of learning tasks>.

[0139] In addition, Figures 11-13 The example given is the task processing unit 121 performing multi-task learning, but the same applies to the task processing unit 121 performing meta-learning.

[0140] Here, as Figure 14A As shown, in existing technologies, contribution is calculated for each piece of data. In contrast, as... Figure 14B As shown, in the data analysis apparatus 12 of Embodiment 1, the contribution is calculated for each learning task that has multiple data.

[0141] Therefore, in the data analysis device 12 of Embodiment 1, it is possible to calculate the contribution of a task that is a collection of multiple data points based on a certain benchmark, rather than the contribution of a single data point. It is possible to provide more intuitive and easier-to-understand explanations for the user, and to determine the reliability based on whether tasks similar to the task that the device wants to perform are suggested as tasks with high contribution.

[0142] Furthermore, the above description uses the case where the information processing system 1 is a text and image classification system as an example, but the information processing system 1, which is the application data analysis device 12, is not limited to this.

[0143] Information processing system 1 may also be, for example, a four-wheel torque control device, a power consumption prediction device, or a production line monitoring system.

[0144] The four-wheel torque control device is a device that receives data representing sensor values ​​from an external source, identifies the road surface condition based on the sensor values ​​in the information processing unit 13, and adjusts the torque output of each wheel of the vehicle.

[0145] In the data analysis of this four-wheel torque control device, learning tasks are defined according to each road surface condition, such as normal road surface, icy and snowy road surface, and gravel road surface. The four-wheel torque control device learns the efficient torque distribution under each road surface condition.

[0146] Furthermore, when switching control modes, the four-wheel torque control device displays information on the dashboard about similar road conditions experienced during the learning process.

[0147] In the prior art, only sensor values ​​used for learning can be displayed. In contrast, in the four-wheel torque control device that applies the data analysis device 12 of Embodiment 1, it is possible to display values ​​on a task-by-task basis using abstract strings or icons such as "icy road surface" or "gravel road surface".

[0148] In addition, the power consumption prediction device is a device that is applied to the operating environment of an air conditioning system or a production line, etc., and that receives data representing sensor values ​​from an external source and predicts power consumption in the information processing unit 13.

[0149] In the data analysis of this power consumption prediction device, learning tasks are defined according to each condition such as usage environment, date, and season. The power consumption prediction device learns future power consumption based on time series data of power usage status under each condition.

[0150] Furthermore, the power consumption prediction device will notify the user whether similar conditions were experienced during the learning process when the device is first used.

[0151] In the prior art, only sensor values ​​used for learning can be displayed. In contrast, in the power consumption prediction device that applies the data analysis device 12 of Embodiment 1, category-based experience content such as season or type of equipment used can be suggested on a task-by-task basis.

[0152] In addition, the production line monitoring system is a system used in environments where humans and robots work in close proximity, which receives data representing sensor values ​​from the outside and predicts human behavior patterns in the information processing unit 13, and stops the robot's movement in cases where there is a risk of contact.

[0153] In the data analysis of this production line monitoring system, learning tasks are defined according to the environment of each production line. In each environment, the production line monitoring system learns the probability of human intrusion into the area or contact with the robot in the next few seconds based on camera images or LiDAR data.

[0154] Furthermore, when the production line monitoring system is introduced, it will notify the user of data regarding whether the user has experienced a similar production environment during the learning process.

[0155] In the prior art, only sensor values ​​used for learning can be displayed. In contrast, in the production line monitoring system that applies the data analysis device 12 of Embodiment 1, it is possible to provide category-based experience information such as intensity or type of production line on a task-by-task basis.

[0156] Furthermore, the above description uses the example of the task retrieval unit 122 calculating the contribution of the learning task to the output of the final processing in the task processing unit 121 and using it as output data. However, it is not limited to this; the task retrieval unit 122 may also calculate the contribution of the learning task to the output of intermediate processing in the task processing unit 121 and use it as output data.

[0157] For example, when the task processing unit 121 performs learning processing, supplementary adjustment processing, and inference processing, the task retrieval unit 122 can calculate the contribution of the learning task to the learning processing and use it as output data, or it can calculate the contribution of the learning task to the supplementary adjustment processing and use it as output data. Thus, the data analysis device 12 can confirm the state of the information processing system 1 in more detail.

[0158] As described above, according to this embodiment 1, the data analysis apparatus 12 includes: a task processing unit 121, which has a learning processing unit 1211, which takes into input a dataset having multiple learning tasks, each having multiple data, and outputs data representing inference model parameters or metadata; and a task retrieval unit 122, which takes into input a dataset and data representing the output result of the task processing unit 121, and outputs data representing the contribution of the learning tasks in the dataset to the output result, or data representing the learning tasks with characteristics of the contribution.

[0159] Therefore, the data analysis device 12 of Embodiment 1 can analyze data on a task-by-task basis. That is, in the data analysis device 12 of Embodiment 1, explanatory information can be provided on a task-by-task basis, thereby enabling more appropriate reliability evaluation.

[0160] Furthermore, according to Embodiment 1, the learning processing unit 1211 outputs data representing inference model parameters, and the task processing unit 121 has an inference processing unit 1212, which inputs data representing the output result of the learning processing unit 1211 and data representing the inference task, and outputs data representing the inference processing result relative to the inference task or data representing the performance evaluation value of the inference processing result.

[0161] Furthermore, according to Embodiment 1, the task processing unit 121 has an adjustment processing unit 1213, which takes in data representing the output result of the learning processing unit 1211 and data representing the reasoning task, and outputs data representing the parameters of the reasoning model.

[0162] Furthermore, according to Embodiment 1, the task processing unit 121 includes: an adjustment processing unit 1213, which takes in data representing the output result of the learning processing unit 1211 and data representing the reasoning task, and outputs data representing the parameters of the reasoning model; and a reasoning processing unit 1212, which takes in data representing the output result of the adjustment processing unit 1213 and data representing the reasoning task, and outputs data representing the reasoning processing result relative to the reasoning task or data representing the performance evaluation value of the reasoning processing result.

[0163] As described above, the data analysis device 12 of Embodiment 1 is able to analyze data on a task-by-task basis.

[0164] Furthermore, according to Embodiment 1, the task retrieval unit 122 calculates the contribution of the learning task in the dataset to the output of the learning processing unit 1211, i.e., the first contribution, and the contribution of the output of the learning processing unit 1211 to the output of the reasoning processing unit 1212, i.e., the second contribution, and synthesizes the first contribution and the second contribution to calculate the contribution of the learning task to the output of the reasoning processing unit 1212.

[0165] Furthermore, according to Embodiment 1, the task retrieval unit 122 calculates the contribution of the learning task in the dataset to the output of the learning processing unit 1211, i.e., the first contribution, and the contribution of the output of the learning processing unit 1211 to the output of the adjustment processing unit 1213, i.e., the third contribution, and synthesizes the first contribution and the third contribution to calculate the contribution of the learning task to the output of the adjustment processing unit 1213.

[0166] Furthermore, according to Embodiment 1, the task retrieval unit 122 calculates the contribution of the learning task in the dataset to the output result of the learning processing unit 1211 (i.e., the first contribution), the contribution of the output result of the learning processing unit 1211 to the output result of the adjustment processing unit 1213 (i.e., the third contribution), and the contribution of the output result of the adjustment processing unit 1213 to the output result of the reasoning processing unit 1212 (i.e., the fourth contribution), and synthesizes the first contribution, the third contribution, and the fourth contribution to calculate the contribution of the learning task to the output result of the reasoning processing unit 1212.

[0167] As described above, the data analysis device 12 of Embodiment 1 is able to analyze data on a task-by-task basis.

[0168] Furthermore, according to Embodiment 1, the task processing unit 121 determines the inference model parameters based on the overall loss function of the dataset obtained by synthesizing the loss functions of each learning task in the dataset according to each learning task. The task retrieval unit 122 calculates the differential coefficient with respect to the perturbation parameter of the overall dataset loss function based on the loss function of any learning task transformed by the perturbation parameter common to the overall dataset within the learning task, and uses the differential coefficient as the contribution of the learning task.

[0169] Furthermore, according to Embodiment 1, for each process performed by the task retrieval unit 122, the differential coefficients are calculated as the contribution between the input and output in that process and a matrix is ​​formed. The matrix product of the matrices of the multiple processes is then calculated to synthesize the contribution.

[0170] Furthermore, according to Embodiment 1, the task processing unit 121 determines the extreme value of the loss function that realizes the entire data set or the model parameters that are approximations of that value, and the task retrieval unit 122 calculates the contribution of the learning task to the inference model parameters determined by the task processing unit 121 through implicit function differentiation.

[0171] As described above, the data analysis device 12 of Embodiment 1 is able to analyze data on a task-by-task basis.

[0172] Furthermore, according to Embodiment 1, the information processing system 1 includes: a dataset acquisition unit 11, which acquires a dataset having multiple learning tasks, each having multiple data; a data analysis device 12; and an information processing unit 13, which performs information processing on a task-by-task basis based on the output of the data analysis device 12, wherein the data analysis device 12 is input with the dataset acquired by the dataset acquisition unit 11.

[0173] Thus, the information processing system 1 of embodiment 1 is able to process information on a task-by-task basis.

[0174] Furthermore, according to Embodiment 1, the data analysis method includes the following steps: a task processing unit 121 inputs a dataset containing multiple learning tasks, each having multiple data points, and outputs data representing inference model parameters or metadata; and a task retrieval unit 122 inputs a dataset and data representing the output result of the task processing unit 121, and outputs data representing the contribution of the learning tasks in the dataset to the output result, or data representing the learning tasks with characteristics of the contribution.

[0175] Therefore, the data analysis method of Implementation 1 can analyze data on a task-by-task basis.

[0176] Implementation method 2.

[0177] In the data analysis apparatus 12 of Embodiment 2, a case is shown where a dataset is processed using a learning task that has characteristics based on contribution.

[0178] Figure 15 This is a diagram illustrating a structural example of the data analysis device 12 in Embodiment 2. Figure 15 In the data analysis device 12 of Embodiment 2 shown, relative to Figure 2 The data analysis apparatus 12 of Embodiment 1 shown here is supplemented with a data processing unit 123 and a second task processing unit 124. Regarding... Figure 15 Other structural examples in the data analysis device 12 of Embodiment 2 shown are the same as those in Embodiment 1, and are described using the same reference numerals.

[0179] In addition, the task retrieval unit 122 outputs data of learning tasks that represent characteristics of contribution to the data processing unit 123.

[0180] The data processing unit 123 processes the data in the dataset based on the output of the task retrieval unit 122.

[0181] At this point, for example, the data processing unit 123 may also exclude learning tasks that contribute little (e.g., close to 0) to the learning processing results from the learning tasks in the dataset. That is, such learning tasks are considered not to affect inference performance, and therefore are excluded from the dataset in order to reduce the computational cost of calculating explanatory information.

[0182] Alternatively, for example, the data processing unit 123 may add copies of learning tasks that contribute significantly to the inference processing results or performance evaluation values ​​on the positive side to the learning tasks in the dataset. That is, such learning tasks are considered to improve inference performance, and therefore copies of the same learning tasks are added to the dataset.

[0183] Alternatively, for example, the data processing unit 123 may exclude learning tasks from the dataset that have a large negative contribution to the inference processing results or performance evaluation values. That is, such learning tasks are considered to degrade inference performance and are therefore excluded from the dataset.

[0184] For example, such as Figure 15 As shown, the second task processing unit 124 includes a learning processing unit 1241 and a reasoning processing unit 1242 for performing reasoning processing. Figure 15 Multi-task learning is performed in the second task processing unit 124 shown.

[0185] The learning processing unit 1241 takes in the dataset processed by the data processing unit 123, performs learning processing based on the dataset, and outputs data representing the parameters of the inference model.

[0186] The inference processing unit 1242 takes in data representing the output result of the learning processing unit 1241 and data representing the inference task, performs inference processing based on the data representing the output result and the inference data, and outputs data representing the inference processing result relative to the inference task or data representing the performance evaluation value of the inference processing result.

[0187] That is, the second task processing unit 124 has the same functions as the task processing unit 121.

[0188] In addition, Figure 15 The diagram shows the learning and reasoning processes performed by the task processing unit 121 and the second task processing unit 124.

[0189] However, it is not limited to this, and with Figures 3-5 Similarly, the task processing unit 121 and the second task processing unit 124 can perform learning processing, supplementary adjustment processing, and reasoning processing, or they can perform only learning processing, or they can perform both learning processing and supplementary adjustment processing. In this case, multi-task learning or meta-learning is performed in the second task processing unit 124.

[0190] In addition, the above shows the situation where the task retrieval unit 122, the data processing unit 123, and the second task processing unit 124 each perform one processing operation.

[0191] However, it is not limited to this. It can also be configured such that the output of the second task processing unit 124 is input back into the task retrieval unit 122, forming a loop between the task retrieval unit 122, the data processing unit 123 and the second task processing unit 124, and the task retrieval unit 122, the data processing unit 123 and the second task processing unit 124 perform multiple processing operations respectively.

[0192] In addition, the data analysis device 12 of Embodiment 2 can also prompt the user with data representing the selection results of the learning task in the task retrieval unit 122, and ask the user to make a decision on whether to perform data processing and reprocessing.

[0193] Furthermore, in the data analysis apparatus 12 of Embodiment 2, a case is shown where, after the data contained in the dataset is processed in the data processing unit 123, the processed dataset is then reprocessed.

[0194] However, it is not limited to this; the improved dataset, which is the output of the data processing unit 123, can also be output to the outside, thus serving as a data cleaning device. That is, in this case, the second task processing unit 124 is not required.

[0195] In addition, as an effective use of the data analysis device 12 in this embodiment 2, the following methods can be considered, for example.

[0196] For example, it is sometimes impossible to predict in advance which learning task will have a learning effect. Therefore, the following effective utilization method can be considered: the initial dataset input to the data analysis device 12 contains redundant learning tasks in advance, and the data processing unit 123 processes the dataset according to the output of the task retrieval unit 122 in order to filter out data that has a learning effect.

[0197] Additionally, for example, in the case of a classification problem, for a specific learning task in the initial dataset, a new learning task can be created that reduces the number of categories or the number of data points per category and added to the dataset, and the added dataset can be input into the data analysis device 12.

[0198] In addition, for example, besides the learning task that uses the original sensor values, a new learning task consisting of filtered data can be created and added to the initial dataset in advance, and the added dataset can be input into the data analysis device 12.

[0199] In addition, Figure 15 The image shows a case where a second task processing unit 124 is separately provided from the task processing unit 121 in the data analysis device 12.

[0200] However, it is not limited to this, for example, such as Figure 16 As shown, it can also be configured such that the second task processing unit 124 is not provided, and the task processing unit 121 includes the functions of the second task processing unit 124.

[0201] That is, in this case, the learning processing unit 121 of the task processing unit 121 takes the dataset processed by the data processing unit 123 as input and processes it again.

[0202] As described above, according to this embodiment 2, the task retrieval unit 122 outputs data of learning tasks that represent characteristics of contribution, and the data analysis device 12 includes a data processing unit 123, which processes the data in the dataset based on the output of the task retrieval unit 122.

[0203] In addition, according to Embodiment 2, the data analysis device 12 includes a second task processing unit 124, which has a learning processing unit 1241. The learning processing unit 1241 takes in a dataset processed by the data processing unit 123 and outputs data representing inference model parameters or metadata.

[0204] In addition, according to Embodiment 2, the learning processing unit 1211 of the task processing unit 121 inputs the dataset processed by the data processing unit 123 and processes it again.

[0205] Therefore, the data analysis apparatus 12 of Embodiment 2 can further improve reliability compared to the data analysis apparatus 12 of Embodiment 1. That is, in the data analysis apparatus 12 of Embodiment 2, data is processed on a learning task basis, thus enabling efficient data reconstruction operations. In addition, even when it is impossible to predict which learning task will have a learning effect, it is possible to use the following method: perform initial task processing using a dataset with a redundant structure, and then select the desired learning task.

[0206] Implementation method 3.

[0207] In the data analysis apparatus 12 of Embodiment 3, a case is shown where attribute information is attached to a learning task, and the attribute information attached to the learning task with contribution characteristics is extracted.

[0208] Figure 17 This is a diagram illustrating a structural example of the data analysis device 12 in Embodiment 3. Figure 17 In the data analysis device 12 of Embodiment 3 shown, relative to Figure 2 The data analysis apparatus 12 of Embodiment 1 shown here is supplemented with a storage unit 125, a storage processing unit 126, an attribute information extraction unit 127, and an additional information receiving unit 128. Regarding... Figure 17 Other structural examples in the data analysis device 12 of Embodiment 3 shown are the same as those in Embodiment 1, and are described using the same reference numerals.

[0209] In addition, the dataset input into the data analysis device 12 includes attribute information representing the attributes of each learning task. This attribute information may include, for example, a description of the learning task, the date and time the data for that task was acquired, or the conditions under which the data was acquired.

[0210] In addition, the task retrieval unit 122 outputs data of learning tasks that represent characteristics of contribution to the attribute information extraction unit 127.

[0211] The storage unit 125 stores the attribute information attached to the learning task together with the information representing the learning task, according to the processing of the storage processing unit 126.

[0212] The storage unit 125 may be, for example, a non-volatile or volatile semiconductor memory such as RAM (Random Access Memory), ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable ROM), EEPROM (Electrically Powered EPROM), disk, floppy disk, optical disk, compact disk, mini disk, or DVD (Digital Versatile Disc).

[0213] In addition, Figure 17 The diagram shows the storage unit 125 located inside the data analysis device 12. However, it is not limited to this; the storage unit 125 may also be located outside the data analysis device 12.

[0214] Based on the dataset, the storage processing unit 126 establishes an association between information representing the learning task of the dataset and attribute information associated with the learning task, and stores them in the storage unit 125.

[0215] exist Figure 17 The storage unit 125 shown stores information such as "Plants in location 00. Photographed in 19××," which indicates the description of learning task 1 and the date and time of data acquisition for learning task 1, as attribute information relative to learning task 1. Additionally, in Figure 17 The storage unit 125 shown stores information such as "Shot at 00:00 in 19××. Equipment malfunction," indicating the date, time, and conditions of data acquisition for learning task 2, serving as attribute information relative to learning task 2. Additionally, in Figure 17 The storage unit 125 shown stores information such as "Shot in 20×× at 00. Black and white", which indicates the date and time of data acquisition and acquisition conditions of the learning task M, as attribute information relative to the learning task M.

[0216] The attribute information extraction unit 127 extracts attribute information associated with the learning task from the storage unit 125 based on the learning task represented by the data output by the task retrieval unit 122. The data representing the attribute information extracted by the attribute information extraction unit 127 is output to the outside.

[0217] Therefore, the data analysis device 12 can output the attribute information associated with the learning task along with the data representing the learning task output by the task retrieval unit 122 to the outside, thus providing the user with more detailed information.

[0218] The supplementary information receiving unit 128 receives supplementary information input by the user, indicating information about the learning task to be added, as well as attribute information associated with the learning task. For example, supplementary information may include comments left by the user when using the data analysis device 12.

[0219] Then, based on the information received by the additional information receiving unit 128, the storage processing unit 126 adds the additional information to the attribute information attached to the learning task that is the object of the addition and stores the attribute information in the storage unit 125.

[0220] In addition, Figure 17 The diagram shows a case where an additional information receiving unit 128 is provided in the data analysis device 12. However, the additional information receiving unit 128 is not a necessary structure for the data analysis device 12, and it may not be provided in the data analysis device 12.

[0221] In addition, Figure 17The diagram shows a case where, compared to the data analysis apparatus 12 of Embodiment 1, a storage unit 125, a storage processing unit 126, an attribute information extraction unit 127, and an additional information receiving unit 128 are added.

[0222] However, it is not limited to this. Alternatively, a storage unit 125, a storage processing unit 126, an attribute information extraction unit 127, and an additional information receiving unit 128 can be added to the data analysis device 12 of Embodiment 2 to achieve the same effect as described above.

[0223] As described above, according to this embodiment 3, in the dataset, each learning task is accompanied by attribute information representing the attributes of that learning task. The task retrieval unit 122 outputs data representing learning tasks with characteristics of contribution. The data analysis device 12 includes: a storage processing unit 126, which establishes an association between information representing the learning tasks in the dataset and the attribute information accompanying the learning tasks based on the dataset and stores it in a storage unit 125; and an attribute information extraction unit 127, which extracts the attribute information associated with the learning tasks from the storage unit 125 based on the learning tasks represented by the data output by the task retrieval unit 122.

[0224] Therefore, the data analysis device 12 of embodiment 3 can provide users with more detailed information.

[0225] Furthermore, according to this embodiment 3, the data analysis device 12 includes an additional information receiving unit 128, which receives additional information input by the user, indicating a learning task to be added and attribute information associated with the learning task. The storage processing unit 126 adds the additional information to the attribute information associated with the learning task to be added based on the information received by the additional information receiving unit 128 and stores the attribute information in the storage unit 125.

[0226] Therefore, the data analysis device 12 of Embodiment 3 can add information input by the user as attribute information.

[0227] Finally, examples of the hardware structure of the data analysis apparatus 12 in embodiments 1 to 3 will be described with reference to FIG18. Here, an example of the hardware structure of the data analysis apparatus 12 in embodiment 1 is shown, but the hardware structure examples of the data analysis apparatus 12 in embodiments 2 and 3 are the same.

[0228] The functions of the task processing unit 121 and the task retrieval unit 122 in the data analysis device 12 are implemented by the processing circuit 51. The processing circuit 51 can be used as follows: Figure 18A As shown, this is dedicated hardware, or it can be like... Figure 18BThe CPU (also known as Central Processing Unit, processing unit, arithmetic unit, microprocessor, microcomputer, processor, or DSP (Digital Signal Processor)) 52 shown is the CPU that executes the program stored in memory 53.

[0229] When the processing circuit 51 is dedicated hardware, it can be, for example, a single circuit, a composite circuit, a programmable processor, a parallel programmable processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or a combination thereof. The functions of each part of the task processing unit 121 and the task retrieval unit 122 can be implemented separately by the processing circuit 51, or the functions of each part can be uniformly implemented by the processing circuit 51.

[0230] When the processing circuit 51 is a CPU 52, the functions of the task processing unit 121 and the task retrieval unit 122 are implemented through software, firmware, or a combination of software and firmware. The software and firmware are described as programs and stored in the memory 53. The processing circuit 51 implements the functions of each unit by reading and executing the programs stored in the memory 53. That is, the data analysis device 12 includes a memory for storing data that, when executed by the processing circuit 51, will ultimately be executed, for example... Figure 4 The procedure for each step shown. Alternatively, the above procedure can also be described as a procedure for the computer to execute the steps and methods of the task processing unit 121 and the task retrieval unit 122. Here, the memory 53 may be, for example, a non-volatile or volatile semiconductor memory such as RAM, ROM, flash memory, EPROM, EEPROM, a disk, floppy disk, optical disk, compact disk, mini disk, or DVD.

[0231] Furthermore, the functions of the task processing unit 121 and the task retrieval unit 122 can be partially implemented by dedicated hardware and partially by software or firmware. For example, the task processing unit 121 can be implemented by a processing circuit 51 as dedicated hardware, and the task retrieval unit 122 can be implemented by the processing circuit 51 reading and executing the program stored in the memory 53.

[0232] Thus, the processing circuit 51 can implement the above-mentioned functions through hardware, software, firmware, or a combination thereof.

[0233] Furthermore, it is possible to freely combine the various embodiments, or modify any constituent elements of the various embodiments, or omit any constituent elements in each embodiment.

[0234] Industrial availability

[0235] The data analysis device 12 disclosed herein is capable of analyzing data on a task-by-task basis and is suitable for use as a data analysis device 12 for analyzing data.

[0236] Explanation of reference numerals in the attached figures

[0237] 1…Information processing system; 11…Data set acquisition unit; 12…Data analysis device; 13…Information processing unit; 51…Processing circuit; 52…CPU; 53…Memory; 121…Task processing unit; 122…Task retrieval unit; 123…Data processing unit; 124…Second task processing unit; 125…Storage unit; 126…Storage processing unit; 127…Attribute information extraction unit; 128…Additional information receiving unit; 1211…Learning processing unit; 1212…Inference processing unit; 1213…Adjustment processing unit; 1241…Learning processing unit; 1242…Inference processing unit.

Claims

1. A data analysis device, characterized in that, have: The task processing unit has a learning processing unit that takes into input a dataset containing multiple learning tasks, each with multiple data points, and outputs data representing inference model parameters or metadata. and The task retrieval unit takes the dataset and data representing the output result of the task processing unit as input, and outputs data representing the contribution of the learning tasks in the dataset to the output result, or data representing the learning tasks with characteristics of the contribution.

2. The data analysis device according to claim 1, characterized in that, The learning processing unit outputs data representing the parameters of the inference model. The task processing unit has a reasoning processing unit that takes in data representing the output result of the learning processing unit and data representing the reasoning task, and outputs data representing the reasoning processing result relative to the reasoning task or data representing the performance evaluation value of the reasoning processing result.

3. The data analysis device according to claim 1, characterized in that, The task processing unit has an adjustment processing unit that takes into input data representing the output result of the learning processing unit and data representing the inference task, and outputs data representing the parameters of the inference model.

4. The data analysis device according to claim 1, characterized in that, The task processing unit has: The adjustment processing unit takes as input data representing the output result of the learning processing unit and data representing the inference task, and outputs data representing the parameters of the inference model; and The inference processing unit takes as input data representing the output result of the adjustment processing unit and data representing the inference task, and outputs data representing the inference processing result relative to the inference task or data representing the performance evaluation value of the inference processing result.

5. The data analysis device according to claim 2, characterized in that, The task retrieval unit calculates the contribution of the learning task in the dataset to the output of the learning processing unit (i.e., the first contribution) and the contribution of the output of the learning processing unit to the output of the inference processing unit (i.e., the second contribution), and synthesizes the first contribution and the second contribution to calculate the contribution of the learning task to the output of the inference processing unit.

6. The data analysis device according to claim 3, characterized in that, The task retrieval unit calculates the contribution of the learning task in the dataset to the output of the learning processing unit (i.e., the first contribution) and the contribution of the output of the learning processing unit to the output of the adjustment processing unit (i.e., the third contribution), and synthesizes the first contribution and the third contribution to calculate the contribution of the learning task to the output of the adjustment processing unit.

7. The data analysis device according to claim 4, characterized in that, The task retrieval unit calculates the contribution of the learning task to the output of the learning processing unit (i.e., the first contribution), the contribution of the output of the learning processing unit to the output of the adjustment processing unit (i.e., the third contribution), and the contribution of the output of the adjustment processing unit to the output of the inference processing unit (i.e., the fourth contribution), and synthesizes the first contribution, the third contribution, and the fourth contribution to calculate the contribution of the learning task to the output of the inference processing unit.

8. The data analysis apparatus according to any one of claims 1 to 4, characterized in that, The task processing unit determines the inference model parameters based on the overall loss function of the dataset, which is obtained by synthesizing the loss functions of each learning task in the dataset according to each learning task. The task retrieval unit calculates the differential coefficient of the perturbation parameter for the loss function of the entire dataset, based on the loss function of any learning task transformed by the common perturbation parameter of the data within the learning task, and uses the differential coefficient as the contribution of the learning task.

9. The data analysis apparatus according to any one of claims 5 to 7, characterized in that, The task processing unit determines the inference model parameters based on the overall loss function of the dataset, which is obtained by synthesizing the loss functions of each learning task in the dataset according to each learning task. The task retrieval unit calculates the differential coefficient of the perturbation parameter for the loss function of the entire dataset, based on the loss function of any learning task transformed by the common perturbation parameter of the data within the learning task, and uses the differential coefficient as the contribution of the learning task.

10. The data analysis device according to claim 9, characterized in that, For each process performed by the task retrieval unit, the differential coefficients are calculated as the contribution between the input and output of the process to form a matrix, and the matrix product of the matrices of each of the multiple processes is calculated to synthesize the contribution.

11. The data analysis apparatus according to claim 8 or 9, characterized in that, The task processing unit determines the extreme value of the loss function for the entire dataset, or the model parameters that approximate that value. The task retrieval unit calculates the contribution of the learning task to the inference model parameters determined by the task processing unit through implicit function differentiation.

12. The data analysis apparatus according to any one of claims 1 to 11, characterized in that, The task retrieval unit outputs data representing learning tasks with characteristics indicating contribution. The data analysis device includes a data processing unit that processes the data in the dataset based on the output of the task retrieval unit.

13. The data analysis apparatus according to claim 12, characterized in that, It has a second task processing unit, which has a learning processing unit that takes into input the dataset processed by the data processing unit and outputs data representing inference model parameters or metadata.

14. The data analysis device according to claim 12, characterized in that, The task processing unit takes the dataset processed by the data processing unit as input to the learning processing unit and processes it again.

15. The data analysis apparatus according to any one of claims 1 to 14, characterized in that, In the dataset, each learning task is accompanied by attribute information representing the attributes of that learning task. The task retrieval unit outputs data representing learning tasks with characteristics indicating contribution. The data analysis device includes: The storage processing unit, based on the dataset, establishes an association between information representing the learning task of the dataset and attribute information associated with the learning task, and stores them in the storage unit. and The attribute information extraction unit extracts attribute information associated with the learning task from the storage unit based on the learning task represented by the data output by the task retrieval unit.

16. The data analysis apparatus according to claim 15, characterized in that, It has an additional information receiving department that accepts additional information input by the user, indicating the learning task to be added, as well as the attribute information associated with the learning task. The storage processing unit adds additional information to the attribute information attached to the learning task as the object of addition, based on the information received by the additional information receiving unit, and stores it in the storage unit.

17. An information processing system, characterized in that, have: The dataset acquisition department acquires datasets for multiple learning tasks, each containing multiple data points. The data analysis apparatus according to any one of claims 1 to 16; as well as The information processing unit processes information on a task-by-task basis based on the output of the data analysis device. The data analysis device inputs the dataset obtained by the dataset acquisition unit.

18. A data analysis method, characterized in that, It includes the following steps: The task processing unit takes into input a dataset containing multiple learning tasks, each with multiple data points, and outputs data representing inference model parameters or metadata. and The task retrieval unit inputs the dataset and data representing the output result of the task processing unit, and outputs data representing the contribution of the learning tasks in the dataset to the output result, or data representing the learning tasks with characteristics of the contribution.

Images