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

Abstract

The present invention comprises: a task processing unit (121) including a learning processing unit (1211) that receives input of a data set, which includes a plurality of learning tasks each having a plurality of data items, and that outputs data indicating an inference model parameter or metadata; and a task retrieval unit (122) that receives input of data indicating an output result based on the data set and by the task processing unit (121), and outputs data indicating the degree of contribution of a learning task included in the data set with respect to the output result, or data indicating a learning task for which the degree of contribution is characteristic.

Description

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

[0001] The present disclosure relates to a data analysis device, an information processing system, and a data analysis method for analyzing data.

[0002] Patent Literature 1 discloses a method for analyzing factors behind predictions made by a trained machine learning model, reconstructing a training dataset, and re-training. In particular, Patent Literature 1 shows that a search unit performs a sensitivity analysis of the impact of changes in training data on predictions, a confirmation unit presents data with a high impact and asks the user for a decision, and a configuration unit reconstructs data to create data for re-training.

[0003] Japanese Patent Application Laid-Open No. 2022-131406

[0004] On the other hand, when it is desired to improve the reliability of a device that uses multitask learning or meta-learning, it is desirable to evaluate reliability on a task-by-task basis. However, the conventional technology disclosed in Patent Document 1 is not intended to improve reliability, and even if it is used for that purpose, it is not possible to evaluate on a task-by-task basis, and judgments must be made based on the contribution of individual data.

[0005] The present disclosure has been made to solve the above-mentioned problems, and aims to provide a data analysis device that is capable of analyzing data on a task-by-task basis.

[0006] The data analysis device according to the present disclosure is characterized by comprising a task processing unit having a learning processing unit that inputs a dataset having a plurality of learning tasks, each having a plurality of data, and outputs data indicating inference model parameters or metadata, and a task search unit that inputs data indicating the dataset and the output results by the task processing unit, and outputs data indicating the contribution of the learning tasks in the dataset to the output results, or data indicating learning tasks characterized by the contribution.

[0007] According to the present disclosure, with the above configuration, data can be analyzed on a task-by-task basis.

[0008] 1 is a block diagram showing an example of the configuration of an information processing system including a data analysis apparatus according to a first embodiment. FIG. 2 is a block diagram showing an example of the configuration of the data analysis apparatus according to the first embodiment. FIG. 3 is a block diagram showing another example of the configuration of the data analysis apparatus according to the first embodiment. FIG. 4 is a block diagram showing another example of the configuration of the data analysis apparatus according to the first embodiment. FIG. 5 is a block diagram showing another example of the configuration of the data analysis apparatus according to the first embodiment. FIG. 6 is a flowchart showing an example of the operation of the data analysis apparatus according to the first embodiment. FIG. 7 is a diagram showing an example of a dataset and an inference task (test task) used in the data analysis apparatus according to the first embodiment. FIG. 8 is a diagram for explaining an example of the operation of the task processing unit in the case of multi-task learning in the first embodiment. FIG. 9 is a diagram for explaining an example of the operation of the task processing unit in the case of meta-learning in the first embodiment. FIG. 10 is a diagram for explaining an example of calculation of the contribution of a learning task to a learning processing result of the task search unit in the first embodiment. FIG. 11 is a diagram for explaining an example of calculation of the contribution of a learning task to a result of an inference processing of the task search unit in the first embodiment. FIG. 12 is a diagram for explaining an example of calculation of the contribution of a learning task to a result of an additional adjustment processing of the task search unit in the first embodiment. FIG. 13 is a diagram for explaining an example of calculation of the contribution of a learning task to a result of an inference processing after an additional adjustment processing of the task search unit in the first embodiment has been performed. 14A and 14B are diagrams for explaining an example of the operation of the data analysis apparatus according to embodiment 1, where FIG. 14A is a diagram showing the case of the conventional technology and FIG. 14B is a diagram showing the case of the data analysis apparatus according to embodiment 1. A block diagram showing an example of the configuration of a data analysis apparatus according to embodiment 2. A block diagram showing another example of the configuration of the data analysis apparatus according to embodiment 2. A block diagram showing an example of the configuration of a data analysis apparatus according to embodiment 3. FIGS. 18A and 18B are block diagrams showing example hardware configurations of the data analysis apparatus according to embodiments 1 to 3.

[0009] Hereinafter, embodiments will be described in detail with reference to the drawings. Embodiment 1. Fig. 1 is a block diagram showing an example configuration of an information processing system 1 including a data analysis device 12 according to embodiment 1. As shown in Fig. 1, the information processing system 1 includes, for example, a data set acquisition unit 11, a data analysis device 12, and an information processing unit 13. An example of this information processing system 1 is a character image classification system that classifies character images.

[0010] The dataset acquisition unit 11 acquires a dataset. For example, as shown in Fig. 2, the dataset includes a plurality of learning tasks. Each of the learning tasks includes a plurality of data.

[0011] The data analysis device 12 receives the data set acquired by the data set acquisition unit 11 and analyzes the data contained in the data set on a task-by-task basis. An example of the configuration of the data analysis device 12 will be described later.

[0012] 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 character image classification system, the information processing unit 13 classifies character images indicated by data input to the information processing system 1 on a task-by-task basis.

[0013] Next, a configuration example of the data analysis device 12 according to the first embodiment will be described with reference to Fig. 2. The data analysis device 12 includes, for example, a task processing unit 121 and a task search unit 122, as shown in Fig. 2.

[0014] 2, the task processing unit 121 includes a learning processing unit 1211 and an inference processing unit 1212. The task processing unit 121 shown in FIG. 2 performs multitask learning.

[0015] The learning processing unit 1211 receives a data set, performs learning processing based on the data set, and outputs data indicating inference model parameters. In the example of Figure 2, the learning processing unit 1211 receives a data set including learning tasks 1 to M. In the example of Figure 2, learning task 1 includes data 1-1 to 1-3, and learning task M includes data M-1 to M-3.

[0016] The inference processing unit 1212 receives data indicating the output result from the learning processing unit 1211 and data indicating the inference task, performs inference processing based on the data indicating the output result and the inference data, and outputs the inference processing result for the inference task or data indicating the performance evaluation value of the inference processing result. The inference task is a test task or a target task.

[0017] 2 illustrates a case in which the task processing unit 121 performs learning processing and inference processing. However, this is not limiting, and for example, the task processing unit 121 may perform learning processing, then additional adjustment processing, and then inference processing. That is, in this case, as shown in FIG. 3, the task processing unit 121 includes an adjustment processing unit 1213 in addition to a learning processing unit 1211 and an inference processing unit 1212. The task processing unit 121 illustrated in FIG. 3 performs multitask learning or meta-learning.

[0018] In this case, the learning processing unit 1211 receives a data set, performs learning processing based on the data set, and outputs data indicating inference model parameters or data indicating meta-parameters. That is, the learning processing unit 1211 outputs data indicating inference model parameters when performing multi-task learning, and outputs data indicating meta-parameters when performing meta-learning.

[0019] The adjustment processing unit 1213 receives data indicating the output results from the learning processing unit 1211 and data indicating an inference task, performs additional adjustment processing based on the data indicating the output results and the inference task, and outputs data indicating inference model parameters. Here, if the output results from the learning processing unit 1211 are data indicating inference model parameters, the adjustment processing unit 1213 adjusts the inference model parameters in the additional adjustment processing and outputs data indicating the adjusted inference model parameters. Furthermore, if the output results from the learning processing unit 1211 are meta parameters, the adjustment processing unit 1213 determines the inference model parameters in the additional adjustment processing and outputs data indicating the inference model parameters.

[0020] In addition, the inference processing unit 1212 inputs data indicating the output result from the adjustment processing unit 1213 and data indicating the inference task, performs inference processing based on the data indicating the output result and the inference task, and outputs data indicating the inference processing result for the inference task or the performance evaluation value of the inference processing result.

[0021] Furthermore, for example, the task processing unit 121 may perform only the learning process without performing the inference process. That is, in this case, for example, as shown in Fig. 4, the task processing unit 121 has a learning processing unit 1211. The task processing unit 121 shown in Fig. 4 performs multitask learning or meta-learning.

[0022] In this case, the learning processing unit 1211 receives a data set, performs learning processing based on the data set, and outputs data indicating inference model parameters or data indicating meta-parameters. That is, the learning processing unit 1211 outputs data indicating inference model parameters when performing multi-task learning, and outputs data indicating meta-parameters when performing meta-learning.

[0023] Furthermore, for example, the task processing unit 121 may perform only the learning process and the additional adjustment process without performing the inference process. That is, in this case, for example, as shown in Fig. 5, the task processing unit 121 has a learning processing unit 1211 and an adjustment processing unit 1213. The task processing unit 121 shown in Fig. 5 performs multitask learning or meta-learning.

[0024] In this case, the learning processing unit 1211 receives a data set, performs learning processing based on the data set, and outputs data indicating inference model parameters or data indicating meta-parameters. That is, the learning processing unit 1211 outputs data indicating inference model parameters when performing multi-task learning, and outputs data indicating meta-parameters when performing meta-learning.

[0025] The task search unit 122 performs a calculation process to calculate the contribution of the learning tasks included in the dataset to the output result based on the dataset and the output result from the task processing unit 121. Alternatively, the task search unit 122 performs the calculation process and a selection process to select learning tasks with characteristic contributions based on the results of the calculation process. In this case, the task search unit 122 selects at least one of learning tasks with a large contribution to the output result and learning tasks with a small contribution to the output result in the selection process. Data indicating the contribution calculated by the task search unit 122 or data indicating the learning tasks (identification information) selected by the task search unit 122 is output to the outside. This allows the user to understand the contribution or learning tasks with characteristic contributions.

[0026] The "degree of contribution" means, for example, the amount of fluctuation in the calculation process result when the data used in the calculation or the processing method thereof is changed based on some rule in the calculation process of a quantity such as a performance evaluation value or a predicted value, or an approximate value thereof. The amount of fluctuation may be changed to a fluctuation rate if the data is parameterized data.

[0027] For example, the amount of variation may be the amount of variation of the performance index when learning is performed after excluding specific data from the dataset. Furthermore, for example, the amount of variation may be the amount of variation of the performance index when learning is performed after replacing specific data in the dataset with dummy data. Furthermore, for example, the amount of variation may be the differential coefficient of the weight parameter of the performance index when learning is performed using a weighted average of specific data and the data values ​​of dummy data in the dataset.

[0028] Furthermore, the "contribution of each task" refers to, for example, the amount of change in the calculation processing result when some kind of coordinated change is made to the entire data included in each task or to the processing method thereof, or an approximation thereof. Note that the amount of change may be changed to a rate of change if the data is parameterized data.

[0029] For example, the variation may be the variation of the performance index when a specific task is removed from the dataset during learning. Alternatively, the variation may be the variation of the performance index when a specific task in the dataset is replaced with a dummy task during learning. Alternatively, the variation may be the differential coefficient of the performance index with respect to a weight parameter when all data included in the specific task in the dataset is learned using a weighted average of the data values ​​of the dummy data using a common weight parameter.

[0030] In addition, when the task processing unit 121 performs multiple processes, for example, the task search unit 122 calculates the contribution between the input and output of each process and combines the contributions to calculate the contribution of the learning task to the final output result of the task processing unit 121.

[0031] For example, consider a case where the task processing unit 121 performs learning processing and inference processing, and the task search unit 122 calculates the contribution of the learning task to the output result by the inference processing unit 1212. In this case, the task search unit 122 first calculates the contribution of the learning task to the output result by the learning processing unit 1211. This contribution is referred to as the first contribution. The task search unit 122 also calculates the contribution of the output result by the learning processing unit 1211 to the output result by the inference processing unit 1212. This contribution is referred to as the second contribution. The task search unit 122 then calculates the contribution of the learning task to the output result by the inference processing unit 1212 by combining the first contribution and the second contribution.

[0032] Also, for example, consider a case where the task processing unit 121 performs learning processing, additional adjustment processing, and inference processing, and the task searching unit 122 calculates the contribution of the learning task to the output result by the inference processing unit 1212. In this case, the task searching unit 122 first calculates the contribution of the learning task to the output result by the learning processing unit 1211. This contribution is referred to as the first contribution. The task searching unit 122 also calculates the contribution of the output result by the learning processing unit 1211 to the output result by the adjustment processing unit 1213. This contribution is referred to as the third contribution. The task searching unit 122 also calculates the contribution of the output result by the adjustment processing unit 1213 to the output result by the inference processing unit 1212. This contribution is referred to as the fourth contribution. The task searching unit 122 then calculates the contribution of the learning task to the output result by the inference processing unit 1212 by combining the first, third, and fourth contributions.

[0033] Also, for example, consider a case where the task processing unit 121 performs learning processing and adjustment processing, and the task searching unit 122 calculates the contribution of the learning task to the output result by the adjustment processing unit 1213. In this case, the task searching unit 122 first calculates the contribution of the learning task to the output result by the learning processing unit 1211. This contribution is referred to as the first contribution. The task searching unit 122 also calculates the contribution of the output result by the learning processing unit 1211 to the output result by the adjustment processing unit 1213. This contribution is referred to as the third contribution. The task searching unit 122 then calculates the contribution of the learning task to the output result by the adjustment processing unit 1213 by combining the first and third contributions.

[0034] Next, an example of the operation of the data analysis device 12 according to the first embodiment will be described with reference to Fig. 6. Note that, although an example of the operation will be described below when the task processing unit 121 has the configuration shown in Fig. 2, the same applies to example operations when the task processing unit 121 has the configurations shown in Figs. 3 to 5. Also, the following describes a case where the task search unit 122 selects learning tasks with distinctive contribution degrees.

[0035] In an example of operation of the data analysis device 12 according to the first embodiment, the task processing unit 121 performs a learning process and an inference process (step ST101), as shown in FIG. 6 . That is, the learning processing unit 1211 receives a data set having a plurality of learning tasks, each of which has a plurality of data, performs a learning process based on the data set, and outputs data indicating inference model parameters. The inference processing unit 1212 then receives data indicating the output results from the learning processing unit 1211 and data indicating the inference tasks, performs an inference process based on the data indicating the output results and the inference tasks, and outputs an inference process result for the inference task or data indicating a performance evaluation value of the inference process result.

[0036] Next, the task search unit 122 performs a calculation process to calculate the contribution of the learning task included in the data set to the output result based on the data set and the output result from the task processing unit 121 (step ST102).

[0037] In this case, the task search unit 122 first calculates the contribution of the learning task to the output result by the learning processing unit 1211. This contribution is referred to as the first contribution. The task search unit 122 also calculates the contribution of the output result by the learning processing unit 1211 to the output result by the inference processing unit 1212. This contribution is referred to as the second contribution. The task search unit 122 then calculates the contribution of the learning task to the output result by the inference processing unit 1212 by combining the first contribution and the second contribution.

[0038] Next, the task search unit 122 performs a selection process to select learning tasks with characteristic contribution degrees based on the results of the calculation process (step ST103). In this process, the task search unit 122 selects at least one of learning tasks with a large contribution degree to the output result and learning tasks with a small contribution degree to the output result. Data indicating the learning tasks (identification information) selected by the task search unit 122 is output to the outside. This allows the user to understand the learning tasks with the characteristic contribution degrees.

[0039] FIG. 7 is a diagram illustrating an example of a dataset and an inference task (test task) used in the data analysis apparatus 12 according to the first embodiment. FIG. 7 illustrates an example of a dataset for classifying multiple types of character images. Note that in FIG. 7 , if the dataset is a dataset for multitask learning, the learning task included in the dataset includes only the training data shown on the left side of FIG. 7 . Also, in FIG. 7 , if the dataset is a dataset for meta-learning, the learning task included in the dataset includes both the training data shown on the left side of FIG. 7 and the test data shown on the right side. That is, in the case of meta-learning, it is necessary to learn meta-parameters to improve the test performance of the learning results of each task. Therefore, the learning task includes test data in addition to the training data. Furthermore, each learning task classifies images by character type. Furthermore, as shown in FIG. 7 , the tasks used for learning are generally of different types from the target task or test task.

[0040] 8 is a diagram for explaining an example of the operation of the task processing unit 121 in the case of multitask learning according to the first embodiment. The additional adjustment process is generally called "additional learning" or "fine tuning." The processing flow shown in FIG. 8 is from left to right.

[0041] In Fig. 8, the first row shows a case where the task processing unit 121 performs only learning processing, as shown in Fig. 4. In this case, the learning processing unit 1211 inputs a dataset, determines inference model parameters based on the learning data contained in the dataset, and outputs data indicating the inference model parameters.

[0042] The second row shows a case where the task processing unit 121 performs learning processing and inference processing, as shown in Fig. 2, for example. In this case, the learning processing unit 1211 inputs a dataset, determines inference model parameters based on the learning data contained in the dataset, and outputs data indicating the inference model parameters. The inference processing unit 1212 also inputs data indicating the inference model parameters and data indicating an inference task output by the learning processing unit 1211, performs performance evaluation of the inference processing or the inference processing result based on the inference model parameters and the inference task, and outputs data indicating a performance evaluation value of the inference processing or the inference processing result.

[0043] The third row shows a case where the task processing unit 121 performs learning processing and additional adjustment processing, as shown in Figure 5, for example. In this case, the learning processing unit 1211 inputs a dataset, determines inference model parameters based on the learning data contained in the dataset, and outputs data indicating the inference model parameters. Furthermore, the adjustment processing unit 1213 inputs data indicating the inference model parameters output by the learning processing unit 1211 and data indicating an inference task, improves the inference model parameters based on the inference task, and outputs data indicating the inference model parameters.

[0044] The fourth row shows a case where the task processing unit 121 performs learning processing, additional adjustment processing, and inference processing, as shown in FIG. 3 . In this case, the learning processing unit 1211 inputs a dataset, determines inference model parameters based on the learning data contained in the dataset, and outputs data indicating the inference model parameters. The adjustment processing unit 1213 inputs data indicating the inference model parameters and data indicating an inference task output by the learning processing unit 1211, improves the inference model parameters, and outputs data indicating the inference model parameters. The inference processing unit 1212 inputs data indicating the inference model parameters and data indicating an inference task output by the adjustment processing unit 1213, performs performance evaluation of the inference processing or the inference processing result based on the inference model parameters and the inference task, and outputs data indicating a performance evaluation value of the inference processing or the inference processing result.

[0045] FIG. 9 is a diagram illustrating an example of the operation of the task processing unit 121 in the case of meta-learning according to the first embodiment. Generally, the learning process is often called "meta-learning," and the additional adjustment process is often called "adaptation" or "learning." The processing flow shown in FIG. 9 is in order from left to right. In addition, in the case of meta-learning, when performing inference processing, additional adjustment processing is required in the preceding stage, so there is no equivalent to the second stage in the case of multi-task learning shown in FIG. 8.

[0046] 8, the first row shows a case where the task processing unit 121 performs only the learning process, as shown in Fig. 4. In this case, the learning processing unit 1211 inputs a dataset, determines inference model parameters based on the learning data contained in the dataset, determines meta parameters that will provide the best inference performance using test data, and outputs data indicating the meta parameters.

[0047] The third row shows a case where the task processing unit 121 performs learning processing and additional adjustment processing, as shown in Figure 5, for example. In this case, the learning processing unit 1211 inputs a dataset, determines inference model parameters based on the learning data contained in the dataset, determines meta parameters that will optimize inference performance using test data, and outputs data indicating the meta parameters. Furthermore, the adjustment processing unit 1213 inputs data indicating the meta parameters output by the learning processing unit 1211 and data indicating an inference task, determines inference model parameters based on the meta parameters and the inference task, and outputs data indicating the inference model parameters.

[0048] The fourth row shows a case where the task processing unit 121 performs learning processing, additional adjustment processing, and inference processing, as shown in FIG. 3 . In this case, the learning processing unit 1211 inputs a dataset, determines inference model parameters based on the learning data contained in the dataset, determines meta parameters that optimize inference performance using test data, and outputs data indicating the meta parameters. The adjustment processing unit 1213 inputs data indicating the meta parameters and data indicating the inference task output by the learning processing unit 1211, determines inference model parameters based on the meta parameters and the inference task, and outputs data indicating the inference model parameters. The inference processing unit 1212 inputs data indicating the inference model parameters and data indicating the inference task output by the adjustment processing unit 1213, performs performance evaluation of the inference processing or the inference processing results based on the inference model parameters and the inference task, and outputs data indicating a performance evaluation value of the inference processing or the inference processing results.

[0049] 10 is a diagram illustrating an example of the calculation operation of the task search unit 122 in embodiment 1 to calculate the contribution of a learning task to a learning processing result. As shown in Fig. 10, for example, when calculating the contribution of a learning task to a learning processing result, the task search unit 122 calculates a sensitivity matrix A, which is a differential coefficient with respect to the perturbation parameters. In this case, the size of the sensitivity matrix A, which is the contribution calculated by the task search unit 122, is <number of inference model parameters> x <number of learning tasks>.

[0050] That is, for example, the task processing unit 121 determines the inference model parameters based on a loss function for the entire dataset. The loss function for the entire dataset is obtained by combining the loss functions for each learning task in the dataset for each learning task. In this case, the task search unit 122 calculates a derivative of the perturbation parameter common to all data in an arbitrary one learning task based on a loss function obtained by modifying the loss function for the entire dataset with respect to the perturbation parameter, and regards the derivative as the contribution of the learning task.

[0051] Here, the inference model parameters (θ), the dataset with M training tasks (D １ , ..., D Ｍ ), and a loss function (L) for each learning task, consider the perturbation of the loss function when the loss function (C) for all data (x) is expressed as in the following equation (1):

[0052] Then, for this formula (1), consider the loss function (C') when any one of the training tasks (j) in the dataset is perturbed by a perturbation parameter (ε) as shown in the following formula (2):

[0053] Here, the value of θ that minimizes C' depends on ε. Therefore, the task search unit 122 determines the value of the differential coefficient when ε = 0 as the contribution of the training task (j) to the training processing result. The task search unit 122 then performs the above process for all training tasks in the dataset to calculate the contribution of each training task.

[0054] The output of the learning process is expressed as an implicit function of the perturbation parameters, for example, by minimizing the loss function. Therefore, the differential coefficient between variables that gives the solution to the minimization problem can be calculated using implicit function differentiation, as in conventional techniques. That is, for example, the task processing unit 121 determines model parameters that achieve the extreme value of the loss function for the entire data set or approximate values ​​of that value. In this case, the task search unit 122 then calculates the contribution of the learning task to the inference model parameters determined by the task processing unit 121 using implicit function differentiation.

[0055] Here, in the prior art, as in the following formula (3), the situation in which data is distinguished by learning tasks is not taken into consideration, and the perturbation of the loss function is considered for each individual piece of data (x). In contrast, the data analysis device 12 according to the first embodiment considers the perturbation of the loss function for each learning task (D), as in formula (2). In other words, the data analysis device 12 according to the first embodiment bundles multiple pieces of data into a single learning task, and considers the coordinated perturbation of all the data in that task.

[0056] 10 shows an example in which the task search unit 122 calculates the differential coefficient by numerical differentiation. However, the present invention is not limited to this, and the task search unit 122 can also calculate a precise differential coefficient by using various mathematical facts.

[0057] Furthermore, when the task processing unit 121 performs multiple processes, for example, the task search unit 122 calculates a differential coefficient as the contribution between the input and output of each process performed by the task processing unit 121 to construct a matrix, and synthesizes the contribution by calculating the matrix product of the matrices for each of the multiple processes.

[0058] 11, when the task processing unit 121 performs learning processing and inference processing, the task searching unit 122 calculates, in addition to the sensitivity matrix A, a sensitivity matrix B, which is a differential coefficient, as the contribution of the output result by the learning processing unit 1211 to the output result by the inference processing unit 1212. In this case, the size of the sensitivity matrix B, which is the contribution calculated by the task searching unit 122, is <the number of performance evaluation values> × <the number of inference model parameters>.

[0059] 11, the task search unit 122 calculates the contribution of each learning task to the output result by the inference processing unit 1212 by calculating the matrix product (B·A) of the sensitivity matrix A and the sensitivity matrix B. This contribution is a matrix whose size is <number of performance evaluation values> × <number of learning tasks>.

[0060] 12, when the task processing unit 121 performs learning processing and additional adjustment processing, the task searching unit 122 calculates, in addition to the sensitivity matrix A, a sensitivity matrix C, which is a differential coefficient, as the contribution of the output result by the learning processing unit 1211 to the output result by the adjustment processing unit 1213. In this case, the size of the sensitivity matrix C, which is the contribution calculated by the task searching unit 122, is <number of inference model parameters> × <number of inference model parameters>.

[0061] 12, the task search unit 122 calculates the contribution of each learning task to the output result by the adjustment processing unit 1213 by calculating the matrix product (C·A) of the sensitivity matrix A and the sensitivity matrix C. This contribution is a matrix whose size is <number of inference model parameters> × <number of learning tasks>.

[0062] 13, when the task processing unit 121 performs learning processing, additional adjustment processing, and inference processing, the task searching unit 122 calculates a sensitivity matrix B', which is a differential coefficient, as the contribution of the output result by the adjustment processing unit 1213 to the output result by the inference processing unit 1212, in addition to the sensitivity matrix A and the sensitivity matrix C. In this case, the size of the sensitivity matrix B', which is the contribution calculated by the task searching unit 122, is <the number of performance evaluation values> × <the number of inference model parameters>.

[0063] 13, the task search unit 122 calculates the contribution of each learning task to the output result by the inference processing unit 1212 by calculating the matrix product (B' C A) of the sensitivity matrix A, the sensitivity matrix C, and the sensitivity matrix B'. This contribution is a matrix whose size is <number of performance evaluation values> × <number of learning tasks>.

[0064] 11 to 13, the case where the task processing unit 121 performs multitask learning has been described as an example, but the same applies to the case where the task processing unit 121 performs meta-learning.

[0065] Here, as shown in Figure 14A, in the conventional technology, the contribution degree is calculated for each data item. In contrast, as shown in Figure 14B, in data analysis device 12 according to embodiment 1, the contribution degree is calculated for each learning task having multiple data items. As a result, data analysis device 12 according to embodiment 1 can calculate the contribution degree of a task that groups together data items based on a certain criterion rather than for each individual data item, and can present an explanation that is more intuitively understandable to the user. Furthermore, reliability can be determined based on whether tasks with high contribution degrees similar to the task that the user wants the device to perform are presented.

[0066] Although the above description has been given using an example in which the information processing system 1 is a character image classification system, the information processing system 1 to which the data analysis device 12 is applied is not limited to this. The information processing system 1 may be, for example, a four-wheel torque control device, a power consumption prediction device, or a production line monitoring system.

[0067] The four-wheel torque control device is a device that inputs data indicating sensor values ​​from an external source, recognizes road surface conditions from the sensor values ​​in an information processing unit 13, and adjusts the torque output of each wheel of the vehicle. In data analysis in this four-wheel torque control device, learning tasks are defined for each road surface condition, such as a normal road surface, a snowy road, and a gravel road, and the four-wheel torque control device learns efficient torque distribution for each road surface condition. When switching control patterns, the four-wheel torque control device then displays on the dashboard information on similar road surface conditions experienced during learning.

[0068] In contrast to the conventional technology, which can only display the sensor values ​​used in learning, the four-wheel torque control device to which the data analysis device 12 according to the first embodiment is applied can display abstract character strings or icons, such as "snowy road" or "gravel road," for each task.

[0069] The power consumption prediction device is applied to the operating environment of an air conditioning system or a production line, and receives data indicating sensor values ​​from an external device to predict power consumption in an information processing unit 13. In data analysis, the power consumption prediction device defines learning tasks for each condition, such as the operating environment, date, and season, and learns future power consumption from time-series data on the power usage status for each condition. When the device is first used, the power consumption prediction device notifies the user whether similar conditions have been experienced during the learning process.

[0070] In contrast to the conventional technology, which can only display the sensor values ​​used in learning, the power consumption prediction device to which the data analysis device 12 according to the first embodiment is applied can present the experience content based on categories such as the season or the type of equipment used for each task.

[0071] The production line monitoring system is also applicable to environments where humans and robots work closely together, and receives data indicating sensor values ​​from an external source. The information processing unit 13 predicts human behavior patterns and halts robot operation if there is a risk of contact. In data analysis, this production line monitoring system defines a learning task for each production line environment, and the production line monitoring system learns, from camera images or Lidar data, areas where humans may enter or the probability of contact with a robot in the next few seconds. When the production line monitoring system is introduced, it notifies the user whether data from a similar production environment has been experienced during learning.

[0072] While conventional techniques can only display the sensor values ​​used in learning, the production line monitoring system to which the data analysis device 12 according to the first embodiment is applied can present experience content based on categories such as congestion or production line type for each task.

[0073] In the above description, the task search unit 122 calculates the contribution of a learning task to the output result of the final processing in the task processing unit 121 and sets the calculated result as output data. However, this is not limiting. The task search unit 122 may also calculate the contribution of a learning task to the output result of an intermediate processing in the task processing unit 121 and set the calculated result as output data. For example, when the task processing unit 121 performs a learning process, an additional adjustment process, and an inference process, the task search unit 122 may calculate the contribution of a learning task to the learning process and set the calculated result as output data, or may calculate the contribution of a learning task to the additional adjustment process and set the calculated result as output data. This allows the data analysis device 12 to check the status of the information processing system 1 in more detail.

[0074] As described above, according to the first embodiment, the data analysis device 12 includes a task processing unit 121 having a learning processing unit 1211 that inputs a dataset having multiple learning tasks, each having multiple data, and outputs data indicating inference model parameters or metadata, and a task search unit 122 that inputs data indicating the dataset and the output results from the task processing unit 121, and outputs data indicating the contribution of the learning tasks in the dataset to the output results, or data indicating learning tasks with characteristic contributions. This enables the data analysis device 12 according to the first embodiment to analyze data on a task-by-task basis. In other words, the data analysis device 12 according to the first embodiment can present explanatory information on a task-by-task basis, thereby enabling more appropriate evaluation of reliability.

[0075] Moreover, according to the first embodiment, the learning processing unit 1211 outputs data indicating inference model parameters, and the task processing unit 121 has an inference processing unit 1212 that receives data indicating the output result by the learning processing unit 1211 and data indicating the inference task, and outputs data indicating the inference processing result for the inference task or a performance evaluation value of the inference processing result. Also, according to the first embodiment, the task processing unit 121 has an adjustment processing unit 1213 that receives data indicating the output result by the learning processing unit 1211 and data indicating the inference task, and outputs data indicating the inference model parameters. Also, according to the first embodiment, the task processing unit 121 has an adjustment processing unit 1213 that receives data indicating the output result by the learning processing unit 1211 and data indicating the inference task, and outputs data indicating the inference model parameters, and an inference processing unit 1212 that receives data indicating the output result by the adjustment processing unit 1213 and data indicating the inference task, and outputs data indicating the inference processing result for the inference task or a performance evaluation value of the inference processing result. As a result, the data analysis apparatus 12 according to the first embodiment can analyze data on a task-by-task basis.

[0076] Moreover, according to the first embodiment, the task search unit 122 calculates a first contribution which is the contribution of the learning task included in the dataset to the output result by the learning processing unit 1211 and a second contribution which is the contribution of the output result by the learning processing unit 1211 to the output result by the inference processing unit 1212, and combines the first contribution and the second contribution to calculate the contribution of the learning task to the output result by the inference processing unit 1212. Moreover, according to the first embodiment, the task search unit 122 calculates a first contribution which is the contribution of the learning task included in the dataset to the output result by the learning processing unit 1211 and a third contribution which is the contribution of the output result by the learning processing unit 1211 to the output result by the adjustment processing unit 1213, and combines the first contribution and the third contribution to calculate the contribution of the learning task to the output result by the adjustment processing unit 1213. Furthermore, according to the first embodiment, the task search unit 122 calculates a first contribution degree, which is the contribution degree of the learning task of the dataset to the output result by the learning processing unit 1211, a third contribution degree, which is the contribution degree of the output result by the learning processing unit 1211 to the output result by the adjustment processing unit 1213, and a fourth contribution degree, which is the contribution degree of the output result by the adjustment processing unit 1213 to the output result by the inference processing unit 1212, and calculates the contribution degree of the learning task to the output result by the inference processing unit 1212 by combining the first contribution degree, the third contribution degree, and the fourth contribution degree. This enables the data analysis apparatus 12 according to the first embodiment to analyze data on a task-by-task basis.

[0077] According to the first embodiment, the task processing unit 121 determines inference model parameters based on a loss function for the entire dataset obtained by combining the loss functions for each learning task in the dataset for each learning task. The task searching unit 122 calculates a differential coefficient for the perturbation parameter common to all data in the learning task based on a loss function obtained by modifying the loss function for any one learning task with respect to the loss function for the entire dataset. The differential coefficient is used as the contribution of the learning task. According to the first embodiment, the task searching unit 122 calculates a differential coefficient for each process performed by the task processing unit 121 as the contribution between the input and output of the process to construct a matrix, and combines the contributions by calculating the matrix product of the matrices for each of the multiple processes. According to the first embodiment, the task processing unit 121 determines model parameters that achieve an extremum of the loss function for the entire dataset or are approximate to that value. The task searching unit 122 calculates the contribution of the learning task to the inference model parameters determined by the task processing unit 121 using implicit differentiation. As a result, the data analysis apparatus 12 according to the first embodiment can analyze data on a task-by-task basis.

[0078] Furthermore, according to the first embodiment, the information processing system 1 includes a dataset acquisition unit 11 that acquires a dataset having a plurality of learning tasks, each having a plurality of data, a data analysis device 12, and an information processing unit 13 that performs information processing on a task-by-task basis based on the output result of the data analysis device 12, and the data analysis device 12 inputs the dataset acquired by the dataset acquisition unit 11. This enables the information processing system 1 according to the first embodiment to process information on a task-by-task basis.

[0079] Furthermore, according to the first embodiment, the data analysis method includes the steps of: a task processing unit 121 inputting a dataset having a plurality of learning tasks, each having a plurality of data, and outputting data indicating inference model parameters or metadata; and a task search unit 122 inputting data indicating the dataset and the output result by the task processing unit 121, and outputting data indicating the contribution of the learning tasks in the dataset to the output result, or data indicating learning tasks characterized by the contribution. This makes it possible for the data analysis method according to the first embodiment to analyze data on a task-by-task basis.

[0080] Second Embodiment A data analysis device 12 according to a second embodiment processes a data set based on a learning task having a characteristic contribution degree.

[0081] Fig. 15 is a diagram showing an example of the configuration of data analysis apparatus 12 according to embodiment 2. In data analysis apparatus 12 according to embodiment 2 shown in Fig. 15, a data processing unit 123 and a second task processing unit 124 are added to data analysis apparatus 12 according to embodiment 1 shown in Fig. 2. The other example of the configuration of data analysis apparatus 12 according to embodiment 2 shown in Fig. 15 is the same as the example of the configuration of data analysis apparatus 12 according to embodiment 1, and will be described with the same reference numerals assigned.

[0082] The task search unit 122 outputs data indicating learning tasks with characteristic contribution degrees to the data processing unit 123. The data processing unit 123 processes the data in the dataset based on the output result from the task search unit 122.

[0083] At this time, for example, the data processing unit 123 may exclude, from among the learning tasks in the dataset, learning tasks that have a low degree of contribution to the learning processing results (for example, close to 0). In other words, since such learning tasks are not expected to affect the inference performance, they are excluded from the dataset in order to reduce the computational cost of calculating the explanation information.

[0084] Furthermore, for example, the data processing unit 123 may add a copy of a learning task that has a large positive contribution to the inference processing result or the performance evaluation value among the learning tasks in the dataset. In other words, since such a learning task is thought to improve the inference performance, a copy of the same learning task is added to the dataset.

[0085] Furthermore, for example, the data processing unit 123 may exclude, from among the training tasks in the data set, training tasks that have a large negative contribution to the inference processing result or the performance evaluation value. In other words, such training tasks are excluded from the data set because they are thought to be deteriorating the inference performance.

[0086] The second task processing unit 124 includes a learning processing unit 1241 and an inference processing unit 1242 that performs inference processing, as shown in Fig. 15. The second task processing unit 124 shown in Fig. 15 performs multitask learning.

[0087] The learning processing unit 1241 inputs the data set processed by the data processing unit 123, performs learning processing based on the data set, and outputs data indicating the inference model parameters.

[0088] The inference processing unit 1242 receives data indicating the output result from the learning processing unit 1241 and data indicating the inference task, performs inference processing based on the data indicating the output result and the inference data, and outputs the inference processing result for the inference task or data indicating the performance evaluation value of the inference processing result. In other words, the functions of the second task processing unit 124 are the same as the functions of the task processing unit 121.

[0089] 15 illustrates a case in which the task processing unit 121 and the second task processing unit 124 perform learning processing and inference processing. However, this is not limiting, and similarly to FIGS. 3 to 5, the task processing unit 121 and the second task processing unit 124 may perform learning processing, additional adjustment processing, and inference processing, or may perform only learning processing, or may perform learning processing and additional adjustment processing. In this case, the second task processing unit 124 performs multitask learning or meta-learning.

[0090] The above describes a case where the task search unit 122, the data processing unit 123, and the second task processing unit 124 each perform processing once. However, this is not limiting, and the output result from the second task processing unit 124 may be input again to the task search unit 122, forming a loop between the task search unit 122, the data processing unit 123, and the second task processing unit 124, so that the task search unit 122, the data processing unit 123, and the second task processing unit 124 each perform processing multiple times.

[0091] In addition, the data analysis device 12 according to the second embodiment may present the user with data indicating the intermediate selection results of the learning tasks in the task search unit 122, and ask the user to decide whether or not to process and reprocess the data.

[0092] In the data analysis apparatus 12 according to the second embodiment, the data processing unit 123 processes the data contained in the dataset, and then reprocesses the processed dataset. However, the present invention is not limited to this. The improved dataset output by the data processing unit 123 can be output to an external device, thereby enabling the apparatus to be used as a data cleansing apparatus. In other words, in this case, the second task processing unit 124 is not required.

[0093] The data analysis device 12 according to the second embodiment can be utilized in the following ways. For example, it may be impossible to predict in advance what kind of learning task will have a learning effect. Therefore, one possible application is to include redundant learning tasks in the initial data set input to the data analysis device 12, and then process the data set using the data processing unit 123 to narrow down the data to have a learning effect based on the output result of the task search unit 122. For example, in the case of a classification problem, a new learning task may be created for a specific learning task in the initial data set by reducing the number of classes or the number of data per class in the learning task, and added to the data set. The resulting data set may then be input to the data analysis device 12. For example, in addition to a learning task using raw sensor values, a new learning task made of filtered data may be created and added to the initial data set, and the resulting data set may then be input to the data analysis device 12.

[0094] 15 shows a case where the data analysis device 12 is provided with a second task processing unit 124 separate from the task processing unit 121. However, this is not limiting, and for example, as shown in FIG. 16 , the second task processing unit 124 may not be provided, and the task processing unit 121 may be configured to include the functions of the second task processing unit 124. That is, in this case, the learning processing unit 1211 included in the task processing unit 121 receives as input the data set processed by the data processing unit 123 and performs processing again.

[0095] As described above, according to the second embodiment, the task search unit 122 outputs data indicating learning tasks with characteristic contribution levels, and includes a data processing unit 123 that processes the data contained in the dataset based on the output result of the task search unit 122. Also, according to the second embodiment, the data analysis device 12 includes a second task processing unit 124 that includes a learning processing unit 1241 that receives the dataset processed by the data processing unit 123 and outputs data indicating inference model parameters or metadata. Also, according to the second embodiment, the learning processing unit 1211 of the task processing unit 121 receives the dataset processed by the data processing unit 123 and performs processing again. As a result, the data analysis device 12 according to the second embodiment can further improve reliability compared to the data analysis device 12 according to the first embodiment. That is, the data analysis device 12 according to the second embodiment handles data on a learning task-by-learning task basis, thereby enabling more efficient data reconstruction work. Furthermore, even when it is difficult to predict which learning tasks will have a learning effect, it is possible to perform an initial task processing using a redundantly configured dataset and then narrow down the required learning tasks.

[0096] Third Embodiment In a data analysis device 12 according to a third embodiment, attribute information is attached to learning tasks, and a case will be described in which attribute information attached to learning tasks having characteristic contribution degrees is extracted.

[0097] Fig. 17 is a diagram showing an example of the configuration of data analysis apparatus 12 according to embodiment 3. In data analysis apparatus 12 according to embodiment 3 shown in Fig. 17, a storage unit 125, a storage processing unit 126, an attribute information extraction unit 127, and an additional information receiving unit 128 are added to data analysis apparatus 12 according to embodiment 1 shown in Fig. 2. The other components of data analysis apparatus 12 according to embodiment 3 shown in Fig. 17 are the same as those of data analysis apparatus 12 according to embodiment 1, and will be described with the same reference numerals assigned.

[0098] The dataset input to the data analysis device 12 includes attribute information indicating the attributes of each learning task. Examples of the attribute information attached to a learning task include a description of the learning task, the acquisition date and time of the data included in the learning task, or the acquisition conditions of the data. The task search unit 122 also outputs data indicating learning tasks with characteristic contribution levels to the attribute information extraction unit 127.

[0099] The storage unit 125 stores the attribute information attached to the learning task together with information indicating the learning task in accordance with processing by the storage processing unit 126. Examples of the storage unit 125 include non-volatile or volatile semiconductor memory such as RAM (Random Access Memory), ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable ROM), and EEPROM (Electrically EPROM), a magnetic disk, a flexible disk, an optical disk, a compact disk, a minidisk, and a DVD (Digital Versatile Disc).

[0100] 17 illustrates a case where the storage unit 125 is provided inside the data analysis device 12. However, the present invention is not limited to this, and the storage unit 125 may be provided outside the data analysis device 12.

[0101] Based on the dataset, the storage processing unit 126 associates information indicating the learning tasks included in the dataset with attribute information attached to the learning tasks and stores them in the storage unit 125. The storage unit 125 shown in FIG. 17 stores, as attribute information for learning task 1, a description of learning task 1, such as "Plants from XX region. Photographed on 19XX," and information indicating the acquisition date and time of data included in learning task 1. The storage unit 125 shown in FIG. 17 also stores, as attribute information for learning task 2, information indicating the acquisition date and time and acquisition conditions of data included in learning task 2, such as "Photographed on 19XX at XX. Equipment failure." The storage unit 125 shown in FIG. 17 also stores, as attribute information for learning task M, information indicating the acquisition date and acquisition conditions of data included in learning task M, such as "Photographed on 20XX at XX. Monochrome."

[0102] The attribute information extraction unit 127 extracts attribute information linked to the learning task from the storage unit 125, based on the learning task indicated by the data output by the task search unit 122. The data indicating the attribute information extracted by the attribute information extraction unit 127 is output to the outside. This allows the data analysis device 12 to output to the outside the attribute information attached to the learning task together with the data indicating the learning task output by the task search unit 122, thereby making it possible to present more detailed information to the user.

[0103] The additional information receiving unit 128 receives information input by the user indicating the learning task to be added and additional information for the attribute information attached to the learning task. Examples of additional information include comments left by the user when using the data analysis device 12. 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 to be added and stores the additional information in the storage unit 125.

[0104] 17 illustrates a case where the additional information receiving unit 128 is provided in the data analysis device 12. However, the additional information receiving unit 128 is not an essential component of the data analysis device 12, and the additional information receiving unit 128 does not necessarily have to be provided in the data analysis device 12.

[0105] 17 shows a case where a storage unit 125, a storage processing unit 126, an attribute information extraction unit 127, and an additional information receiving unit 128 are added to the data analysis apparatus 12 according to embodiment 1. However, the present invention is not limited to this, and a storage unit 125, a storage processing unit 126, an attribute information extraction unit 127, and an additional information receiving unit 128 may be added to the data analysis apparatus 12 according to embodiment 2, and the same effects as those described above can be obtained.

[0106] As described above, according to the third embodiment, the data set is provided with attribute information indicating the attributes of each learning task, and the task search unit 122 outputs data indicating learning tasks with characteristic contribution levels, and the data analysis apparatus 12 includes a storage processing unit 126 that links information indicating the learning tasks included in the data set with the attribute information attached to the learning tasks based on the data set and stores them in the storage unit 125, and an attribute information extraction unit 127 that extracts the attribute information linked to the learning task from the storage unit 125 based on the learning task indicated by the data output by the task search unit 122. This enables the data analysis apparatus 12 according to the third embodiment to present more detailed information to the user.

[0107] Furthermore, according to the third embodiment, data analysis device 12 includes additional information receiving unit 128 that receives information input by the user indicating the learning task to be added and additional information for the attribute information attached to the learning task, and storage processing unit 126 adds and stores the additional information to the attribute information attached to the learning task to be added, based on the information received by additional information receiving unit 128. This enables data analysis device 12 according to the third embodiment to add information input by the user as attribute information.

[0108] Finally, with reference to FIG. 18 , an example of the hardware configuration of the data analysis apparatus 12 according to the first to third embodiments will be described. Here, an example of the hardware configuration of the data analysis apparatus 12 according to the first embodiment will be described, but the same applies to the example of the hardware configuration of the data analysis apparatus 12 according to the second and third embodiments. The functions of the task processing unit 121 and the task searching unit 122 in the data analysis apparatus 12 are realized by a processing circuit 51. The processing circuit 51 may be dedicated hardware as shown in FIG. 18A , or may be a CPU (also referred to as a central processing unit, processing unit, arithmetic unit, microprocessor, microcomputer, processor, or DSP (Digital Signal Processor)) 52 that executes a program stored in a memory 53 as shown in FIG. 18B .

[0109] When the processing circuitry 51 is dedicated hardware, the processing circuitry 51 may be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or a combination thereof. The functions of the task processing unit 121 and the task searching unit 122 may be realized individually by the processing circuitry 51, or the functions of the respective units may be realized collectively by the processing circuitry 51.

[0110] When the processing circuitry 51 is a CPU 52, the functions of the task processing unit 121 and the task searching unit 122 are realized by software, firmware, or a combination of software and firmware. The software and firmware are written as programs and stored in the memory 53. The processing circuitry 51 realizes 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 programs that, when executed by the processing circuitry 51, result in the execution of, for example, each step shown in FIG. 4 . These programs can also be said to cause a computer to execute the procedures and methods of the task processing unit 121 and the task searching unit 122. Here, examples of the memory 53 include non-volatile or volatile semiconductor memory such as RAM, ROM, flash memory, EPROM, and EEPROM, magnetic disks, flexible disks, optical disks, compact disks, minidisks, and DVDs.

[0111] It is also possible to realize some of the functions of the task processing unit 121 and the task searching unit 122 with dedicated hardware and some with software or firmware. For example, the function of the task processing unit 121 can be realized by the processing circuitry 51 as dedicated hardware, and the function of the task searching unit 122 can be realized by the processing circuitry 51 reading and executing a program stored in the memory 53.

[0112] In this way, the processing circuitry 51 can realize each of the above-described functions by hardware, software, firmware, or a combination of these.

[0113] It should be noted that the embodiments may be freely combined, or any of the components in each embodiment may be modified, or any of the components in each embodiment may be omitted.

[0114] The data analysis device 12 according to the present disclosure is capable of analyzing data on a task-by-task basis and is suitable for use as a data analysis device 12 that analyzes data.

[0115] 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 search unit, 123 Data processing unit, 124 Second task processing unit, 125 Memory unit, 126 Storage processing unit, 127 Attribute information extraction unit, 128 Additional information reception 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 comprising: a task processing unit having an input of a dataset having a plurality of learning tasks, each having a plurality of data, and a learning processing unit which outputs data indicating inference model parameters or metadata; and a task search unit which inputs data indicating the dataset and an output result by the task processing unit, and outputs data indicating the contribution of the learning tasks in the dataset to the output result, or data indicating a learning task having a characteristic contribution degree.

2. The data analysis device described in claim 1, characterized in that the learning processing unit outputs data indicating inference model parameters, and the task processing unit has an inference processing unit that inputs data indicating the output result by the learning processing unit and data indicating an inference task, and outputs data indicating the inference processing result for the inference task or a performance evaluation value of the inference processing result.

3. The data analysis device according to claim 1, characterized in that the task processing unit has an adjustment processing unit that inputs data indicating the output results from the learning processing unit and data indicating an inference task, and outputs data indicating inference model parameters.

4. The data analysis device according to claim 1, characterized in that the task processing unit has: an adjustment processing unit that inputs data indicating the output result by the learning processing unit and data indicating an inference task, and outputs data indicating inference model parameters; and an inference processing unit that inputs data indicating the output result by the adjustment processing unit and data indicating an inference task, and outputs data indicating the inference processing result for the inference task or a performance evaluation value of the inference processing result.

5. The data analysis device according to claim 2, characterized in that the task search unit calculates a first contribution degree which is the contribution degree of the learning task of the dataset to the output result by the learning processing unit, and a second contribution degree which is the contribution degree of the output result by the learning processing unit to the output result by the inference processing unit, and calculates the contribution degree of the learning task to the output result by the inference processing unit by combining the first contribution degree and the second contribution degree.

6. The data analysis device according to claim 3, characterized in that the task search unit calculates a first contribution degree which is the contribution degree of the learning task of the dataset to the output result by the learning processing unit, and a third contribution degree which is the contribution degree of the output result by the learning processing unit to the output result by the adjustment processing unit, and calculates the contribution degree of the learning task to the output result by the adjustment processing unit by combining the first contribution degree and the third contribution degree.

7. The data analysis device according to claim 4, characterized in that the task search unit calculates a first contribution degree which is the contribution degree of the learning task of the dataset to the output result by the learning processing unit, a third contribution degree which is the contribution degree of the output result by the learning processing unit to the output result by the adjustment processing unit, and a fourth contribution degree which is the contribution degree of the output result by the adjustment processing unit to the output result by the inference processing unit, and calculates the contribution degree of the learning task to the output result by the inference processing unit by combining the first contribution degree, the third contribution degree, and the fourth contribution degree.

8. The data analysis device according to any one of claims 1 to 4, characterized in that the task processing unit determines inference model parameters based on a loss function for the entire dataset obtained by synthesizing loss functions for each learning task in the dataset for each learning task, and the task search unit calculates a differential coefficient for the perturbation parameter common to all data in the learning task based on a loss function obtained by transforming the loss function of any one learning task with respect to the loss function for the entire dataset, and regards the differential coefficient as the contribution of the learning task.

9. The data analysis device according to any one of claims 5 to 7, characterized in that the task processing unit determines inference model parameters based on a loss function for the entire dataset obtained by synthesizing loss functions for each learning task in the dataset for each learning task, and the task search unit calculates a differential coefficient for the perturbation parameter based on a loss function obtained by transforming the loss function of any one learning task with a perturbation parameter common to all data in the learning task, with respect to the loss function for the entire dataset, and regards the differential coefficient as the contribution of the learning task.

10. The data analysis device according to claim 9, characterized in that the task search unit calculates a differential coefficient as the contribution between input and output in each process performed by the task processing unit, constructs a matrix, and synthesizes the contribution by calculating the matrix product of the matrices for each of the multiple processes.

11. The data analysis device according to claim 8 or 9, characterized in that the task processing unit determines model parameters that achieve an extreme value of the loss function for the entire data set or that are approximations to that value, and the task search unit calculates the contribution of the learning task to the inference model parameters determined by the task processing unit by implicit function differentiation.

12. A data analysis device as claimed in any one of claims 1 to 11, characterized in that the task search unit outputs data indicating learning tasks having characteristic contribution levels, and is further provided with a data processing unit that processes the data contained in the dataset based on the output results from the task search unit.

13. The data analysis device according to claim 12, further comprising a second task processing unit having a learning processing unit that inputs the data set processed by the data processing unit and outputs data indicating inference model parameters or metadata.

14. The data analysis device according to claim 12, characterized in that the learning processing unit of the task processing unit receives the data set processed by the data processing unit and processes it again.

15. A data analysis device as claimed in any one of claims 1 to 14, characterized in that the data set is provided with attribute information indicating attributes of each learning task, the task search unit outputs data indicating learning tasks having a characteristic contribution level, and the data processing unit links and stores in a memory unit information indicating the learning tasks contained in the data set and the attribute information attached to the learning tasks based on the data set, and an attribute information extraction unit extracts from the memory unit the attribute information linked to the learning task based on the learning task indicated by the data output by the task search unit.

16. A data analysis device as described in claim 15, further comprising an additional information receiving unit that receives information input by a user indicating the learning task to be added and additional information for attribute information attached to the learning task, and the storage processing unit adds and stores the additional information in the memory unit based on the information received by the additional information receiving unit.

17. An information processing system comprising: a dataset acquisition unit that acquires a dataset having a plurality of learning tasks, each having a plurality of data; a data analysis device according to any one of claims 1 to 16; and an information processing unit that performs information processing on a task-by-task basis based on the output results of the data analysis device, wherein the data analysis device inputs the dataset acquired by the dataset acquisition unit.

18. A data analysis method comprising the steps of: a task processing unit inputting a dataset having a plurality of learning tasks, each having a plurality of data, and outputting data indicating inference model parameters or metadata; and a task search unit inputting data indicating the dataset and an output result by the task processing unit, and outputting data indicating the contribution of the learning tasks in the dataset to the output result, or data indicating learning tasks characterized by the contribution.

Images