Information processing device, control method, and control program

Abstract

To provide an information processing device capable of effectively selecting appropriate data for learning.SOLUTION: An information processing device comprises: a plurality of mutually different extraction units, each configured to extract a feature amount from data; and a determination unit configured to determine the feature amount used for selecting the data to be used as data for learning of a model, based on a plurality of feature amounts each extracted from the plurality of extraction units.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an information processing device, a control method, and a control program. [Background technology]

[0002] In recent years, various technologies using machine learning have been developed. For example, a classifier including a trained model trained by machine learning can classify various types of data into multiple classes. The trained model extracts features from the data and classifies the data into one of multiple classes based on the extracted features. For example, in a product manufacturing site, a classifier that classifies input image data into good or bad products can be used for product inspection.

[0003] Generally, the more training data used in machine learning, the better the performance of a classifier. However, annotation, which assigns correct labels to a large amount of training data, requires a great deal of effort and time.

[0004] To solve this problem, there is a technology called active learning, which selects the minimum amount of data that is estimated to have a high learning effect, and then annotates only the selected data to train the model.

[0005] The following prior art is disclosed in the following patent document: A feature vector is calculated from non-correct data. For each selection candidate, the similarity between the feature vector calculated from the non-correct data of the selection candidate and the feature vector calculated from non-correct data other than the selection candidate is calculated. The feature vector is a vector representation of multiple feature quantities calculated from the data. The non-correct data of the selection candidate with the calculated higher similarity is selected as data expected to be effective for learning. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Japanese Patent Publication No. 2022-184272 Summary of the Invention [Problem to be solved by the invention]

[0007] There are various models capable of extracting features from data, such as convolutional neural networks (CNNs), transformers, and large language models (LLMs). The trends and granularity of the features extracted by these models can vary considerably. Therefore, when selecting training data to be used for training a classifier based on features extracted from data, it may be difficult to select appropriate training data depending on the model used to extract the features. The above-mentioned prior art cannot address this issue.

[0008] The present invention has been made to solve the above-mentioned problems, and an object of the present invention is to provide an information processing device, a control method, and a control program that enable efficient selection of appropriate learning data. [Means for solving the problem]

[0009] The above-mentioned problems of the present invention are solved by the following means.

[0010] (1) An information processing device having a plurality of different extraction units that each extract features from data, and a determination unit that determines the features to be used in selecting the data to be used as training data for a model based on the plurality of features extracted by the plurality of extraction units.

[0011] (2) The information processing device described in (1) above, wherein the determination unit clusters the extracted features for each extraction unit, selects the extraction unit that extracts the features whose number of clusters is close to the number of classes of the task executed by the model, and determines the features extracted by the selected extraction unit as the features to be used for selecting the data.

[0012] (3) The information processing device described in (1) above, wherein the determination unit clusters the extracted features for each extraction unit, selects the extraction unit that extracts the features with small variation in each cluster, and determines the features extracted by the selected extraction unit as the features to be used for selecting the data.

[0013] (4) The information processing device described in (1) above, wherein the determination unit clusters the extracted features for each extraction unit, selects the extraction unit that extracts the features with a large distance between clusters, and determines the features extracted by the selected extraction unit as the features to be used for selecting the data.

[0014] (5) The information processing device described in (1) above, wherein the determination unit selects the extraction unit that extracts, from the features extracted from the labeled data, features that are similar to features extracted from the data with the same label, and determines the features extracted by the selected extraction unit as the features to be used for selecting the data.

[0015] (6) The information processing device described in (1) above, wherein the determination unit selects the extraction unit that extracts, from the features extracted from the labeled data, features that are different from the features extracted from the data with different labels, and determines the features extracted by the selected extraction unit as the features to be used for selecting the data.

[0016] (7) The information processing device according to any one of (2) to (6) above, wherein the determination unit selects one of the extraction units.

[0017] (8) An information processing device described in any of (2) to (6) above, wherein the determination unit determines each value obtained by adding up the features extracted by each of the selected multiple extraction units as the feature to be used in selecting the data.

[0018] (9) An information processing device described in any of (2) to (6) above, wherein the determination unit determines each value obtained by merging the features extracted by each of the selected multiple extraction units as the feature to be used in selecting the data.

[0019] (10) The information processing device described in (1) above further comprises a display unit that displays the process of determining the feature by the determination unit, and a user interface that displays the extraction unit in a user-selectable manner, and the determination unit redetermines the feature extracted by the extraction unit selected in the user interface as the feature to be used in selecting the data.

[0020] (11) The information processing device described in (1) above, wherein when the decision unit determines that there is no feature among the extracted multiple features that can be used to select the data to be used as learning data for the model, the decision unit outputs information indicating that there is no feature that can be used to select the data.

[0021] (12) The information processing device described in (1) above, further comprising an extraction unit addition unit that adds to the extraction unit the model trained using the data selected using the determined feature.

[0022] (13) A control method for causing an information processing device to execute the steps of: (a) extracting features from data using a plurality of different extraction units; and (b) determining the features to be used in selecting the data to be used as training data for a model based on the plurality of features extracted by the plurality of extraction units.

[0023] (14) In the control method described in (13) above, in step (b), the extracted features are clustered for each extraction unit, an extraction unit is selected that extracts the features whose number of clusters is close to the number of classes of the task executed by the model, and the features extracted by the selected extraction unit are determined to be the features to be used for selecting the data.

[0024] (15) In the control method described in (13) above, in step (b), the extracted features are clustered for each extraction unit, an extraction unit that extracts features with small variation in each cluster is selected, and the feature extracted by the selected extraction unit is determined to be the feature to be used for selecting the data.

[0025] (16) In the control method described in (13) above, in step (b), the extracted features are clustered for each extraction unit, an extraction unit that extracts the feature with a large distance between clusters is selected, and the feature extracted by the selected extraction unit is determined to be the feature to be used for selecting the data.

[0026] (17) In the control method described in (13) above, in the step (b), the extraction unit that extracts the feature from the labeled data that is similar to the feature extracted from the data that has the same label is selected, and the feature extracted by the selected extraction unit is determined to be the feature to be used for selecting the data.

[0027] (18) In the control method described in (13) above, in the step (b), an extraction unit is selected that extracts features from the labeled data that are different from features extracted from the data with different labels, and the features extracted by the selected extraction unit are determined to be the features to be used for selecting the data.

[0028] (19) The control method according to (14) above, wherein in the step (b), one of the extraction units is selected.

[0029] (20) The control method according to (14) above, wherein in step (b), each value obtained by adding up the features extracted by the selected plurality of feature extraction units is determined as the feature to be used for selecting the data.

[0030] (21) In the control method described in (14) above, in step (b), each value obtained by merging the features extracted by each of the selected multiple extraction units is determined as the feature to be used for selecting the data.

[0031] (22) The control method according to (13) above, further comprising a step (c) of displaying the process of determining the feature in step (b), and a step (d) of displaying the extraction unit by a user interface so that the extraction unit can be selected by a user, wherein in step (b), the feature extracted by the extraction unit selected in the user interface is redetermined as the feature to be used for selecting the data.

[0032] (23) In the control method described in (13) above, when it is determined that there is no feature among the extracted multiple features that can be used to select the data to be used as learning data for the model in step (b), information indicating that there is no feature that can be used to select the data is output.

[0033] (24) The control method described in (13) above, further comprising a step (d) of adding to the extraction unit the model trained using the data selected using the determined feature.

[0034] (25) A control program for causing a computer to execute the control method according to any one of (13) to (24) above. [Effects of the Invention]

[0035] The feature amounts to be used for selecting learning data are determined based on the plurality of feature amounts extracted by the plurality of extraction units, thereby enabling efficient selection of appropriate learning data. [Brief explanation of the drawings]

[0036] Advantages and features provided by one or more embodiments of the present invention will be more fully understood from the following detailed description and the accompanying drawings, which are for purposes of illustration only and are not intended to be limiting. [Figure 1] FIG. 1 is a diagram illustrating a schematic configuration of an information processing system. [Figure 2] FIG. 2 is a block diagram showing a hardware configuration of the information processing system. [Figure 3A] FIG. 10 is an explanatory diagram showing an example of feature amounts displayed in a feature amount space. [Figure 3B] FIG. 10 is an explanatory diagram showing an example of feature amounts displayed in a feature amount space. [Figure 3C] FIG. 10 is an explanatory diagram showing an example of feature amounts displayed in a feature amount space. [Figure 3D] FIG. 10 is an explanatory diagram showing an example of feature amounts displayed in a feature amount space. [Figure 4] 10 is an explanatory diagram for explaining a method for selecting unlabeled image data by a data selection unit. FIG. [Figure 5] 10 is a flowchart showing the operation of the information processing system. DETAILED DESCRIPTION OF THE INVENTION

[0037] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, the scope of the present invention is not limited to the disclosed embodiments. In the description of the drawings, the same elements are denoted by the same reference numerals, and duplicate explanations will be omitted. Furthermore, the dimensional proportions in the drawings are exaggerated for the convenience of explanation and may differ from the actual proportions.

[0038] Fig. 1 is a diagram showing a schematic configuration of an information processing system 10 including a feature determining unit 100 according to an embodiment. Fig. 2 is a block diagram showing a hardware configuration of the information processing system 10. The feature determining unit 100 constitutes an information processing device.

[0039] 2, the information processing system 10 includes a control unit 11, a storage unit 12, a display unit 13, an input unit 14, and a communication unit 15. These components are connected to each other via a bus 16. The information processing system 10 may be configured by a computer.

[0040] The control unit 110 is configured by a CPU (Central Processing Unit) and performs control and calculation processing of each unit of the information processing system 10 according to a program. The function of the control unit 110 will be described later with reference to FIG.

[0041] The storage unit 120 may be configured with a RAM (Random Access Memory), a ROM (Read Only Memory), and a flash memory. The RAM temporarily stores programs and data as a working area for the control unit 110. The ROM stores various programs and data in advance. The flash memory stores various programs and data, including the operating system.

[0042] The display unit 130 is, for example, a liquid crystal display, and displays various types of information.

[0043] The input unit 140 is configured with, for example, a touch panel and various keys. The input unit 140 is used for various operations and inputs. The input unit 140 accepts a correct label when performing annotation. Hereinafter, image data to which a correct label has been assigned by annotation will also be referred to as "labeled image data." Image data to which no correct label has been assigned will also be referred to as "unlabeled image data."

[0044] The communication unit 150 is an interface for communicating with external devices. For communication, a network interface conforming to standards such as Ethernet (registered trademark), SATA, PCI Express, USB, and IEEE1394 may be used. In addition, for communication, various local connection interfaces such as wireless communication interfaces such as Bluetooth (registered trademark) and IEEE802.11 may be used. The communication unit 150 receives data. For simplicity of explanation, the following description will be given taking the case where the data is image data as an example.

[0045] The following describes the functions of the control unit 110. As shown in Fig. 1, the control unit 11 functions as a feature determining unit 100, a data selecting unit 200, and a learning unit 300 by executing a program.

[0046] The feature amount determination unit 100 includes a plurality of feature amount extraction units 111 and a determination unit 112. The plurality of feature amount extraction units 111 are configured from a first feature amount extraction unit, a second feature amount extraction unit, ..., an Nth feature amount extraction unit, where N is an integer equal to or greater than 2. The feature amount extraction units 111 constitute an extraction unit.

[0047] The multiple feature extraction units 111 are different from each other and extract features from image data. Specifically, each feature extraction unit 111 includes a different trained model. The trained model of the feature extraction unit 111 is a trained model trained to extract features from image data. The trained models of the feature extraction unit 111 include CNN-based, Transformer-based, and LLM-based models. The different trained models include different types of trained models. The different trained models further include trained models trained using different pre-training data. CNN-based trained models include LeNet, VGG, ResNet, etc. Transformer-based trained models include ViT, Swin Transformer, etc. LLM models include BERT, GPT-3, etc. Each feature extraction unit 111 can extract features from image data as feature vectors.

[0048] The determination unit 112 determines features to be used for selecting image data to be used as model training data based on the features extracted by the feature extraction units 111. Hereinafter, features to be used for selecting image data to be used as model training data are also referred to as "data selection features." As will be described later, a model is a model that becomes a trained model by being trained by the training unit 300 using training data. The selection of image data to be used as model training data is performed by the data selection unit 200. The trained model can be used to perform a task of classifying image data into one of multiple classes. For example, in response to input image data of a product, the trained model outputs a classification result indicating whether the product is good or defective. In this case, the trained model classifies the image data into either a defective product class or a good product class. Hereinafter, the model trained by the training unit 300 and the trained model trained by the training unit 300 are also simply referred to as "model."

[0049] 3A to 3D are explanatory diagrams showing examples of features displayed in feature space, in which the distinction between classes is expressed by different colors applied to the plots.

[0050] Figure 3A shows an example of features displayed in feature space when feature extraction is ideal. As shown in Figure 3A, it is desirable to form clusters for each class by grouping features of the same class relatively close together. It is also desirable that the distance between the formed clusters is large.

[0051] Figure 3B shows an example of features displayed in feature space when the extracted features are too coarse. It is undesirable to extract features that do not represent the fine differences in the features of each class contained in the image data, as shown in Figure 3B.

[0052] Figure 3C shows an example of features displayed in feature space when the extracted features are too detailed. It is undesirable to extract features that capture too many subtle differences in the features contained in the image data and fail to capture the features specific to the same class, as shown in Figure 3C.

[0053] 3D shows an example of features displayed in the feature space when bias in the features occurs due to the influence of the pre-training data used in training the feature extraction unit 111. It is not desirable for features extracted from image data of the same class to separate and form multiple clusters, as shown in FIG. 3D. It is also not desirable for features extracted from image data of different classes to be grouped together and form a single cluster.

[0054] The determination unit 112 determines the feature for data selection, for example, by the following method. 1. Extracting features from multiple unlabeled image data (1-1) The determination unit 112 clusters features extracted from image data for each of the multiple feature extraction units 111. The determination unit 112 selects a feature extraction unit 111 that extracts features whose number of clusters obtained by clustering is close to the number of classes of the task executed by the model. For example, the determination unit 112 selects one or more feature extraction units 111 that extract features whose number of clusters is closest to the number of classes of the task executed by the model. The determination unit 112 may select a predetermined number of feature extraction units 111 in descending order of the number of clusters whose number is closest to the number of classes of the task executed by the model. The determination unit 112 determines the features extracted by the selected feature extraction units 111 as features for data selection.

[0055] (1-2) The determination unit 112 clusters the features extracted from the image data for each of the multiple feature extraction units 111. The determination unit 112 selects a feature extraction unit 111 that extracts a feature with small variation in each cluster. For example, the determination unit 112 calculates the average value of the standard deviations of all clusters. Then, the determination unit 112 selects one or more feature extraction units 111 that extract a feature with the smallest calculated average value. The determination unit 112 may select a predetermined number of feature extraction units 111 in ascending order of the calculated average value. The determination unit 112 determines the feature extracted by the selected feature extraction unit 111 as the feature for data selection.

[0056] (1-3) The determination unit 112 clusters the features extracted from the image data for each of the multiple feature extraction units 111. The determination unit 112 selects a feature extraction unit 111 that extracts a feature having a large distance between clusters. For example, the determination unit 112 selects one or more feature extraction units 111 that extract a feature having a large average value of Euclidean distance between clusters. The determination unit 112 may select a predetermined number of feature extraction units 111 in descending order of the calculated average value. The determination unit 112 determines the features extracted by the selected feature extraction units 111 as features for data selection. 2. Extracting features from image data with multiple labels (2-1) The determination unit 112 extracts features from multiple labeled image data. The determination unit 112 selects a feature extraction unit 111 that extracts, from the extracted features, a feature that is similar to a feature extracted from image data with the same label. Specifically, the determination unit 112 calculates the average value for all labels of the standard deviations of the features extracted from image data with the same label. The determination unit 112 then selects one or more feature extraction units 111 that extract features with the smaller calculated average value. The determination unit 112 may select a predetermined number of feature extraction units 111 in ascending order of the calculated average value. The determination unit 112 determines the features extracted by the selected feature extraction units 111 as features for data selection.

[0057] (2-2) The determination unit 112 extracts features from multiple labeled image data. The determination unit 112 selects feature extraction units 111 that extract, from among the extracted features, feature quantities extracted from image data with different labels that are different from each other. Specifically, the determination unit 112 calculates, for each label, the average value of feature quantities extracted from image data with the same label. Then, the determination unit 112 selects one or more feature extraction units 111 that extract feature quantities with the largest difference between the maximum and minimum values ​​of the calculated average values. The determination unit 112 may select a predetermined number of feature extraction units 111 in descending order of the calculated difference. The determination unit 112 determines the feature quantities extracted by the selected feature extraction units 111 as features for data selection.

[0058] In this way, the determining unit 112 determines, from among the feature amounts extracted by the plurality of feature amount extracting units 111, a feature amount that is relatively close to an ideally extracted feature amount, as exemplified in FIG. 3A.

[0059] When selecting a plurality of feature extraction units 111, the determination unit 112 may determine the features for data selection as follows. The determination unit 112 determines each value obtained by adding up the features extracted by the plurality of feature extraction units 111 as the features for data selection. The determination unit 112 may also determine each value obtained by weighting the features extracted by the plurality of feature extraction units 111 as the features for data selection. When extracting features from unlabeled image data, the weighted average may be calculated as follows. The selection order in the selection of the feature extraction units 111 in the above-mentioned (1-1) to (1-3) is scored so that the higher the selection order, the higher the score. Then, a weighted average is calculated by weighting the features extracted by each selected feature extraction unit 111 using the score. When extracting features from labeled image data, the weighted average may be calculated as follows. The selection order in the selection of the feature extraction units 111 in the above-mentioned (2-1) to (2-3) is scored so that the higher the selection order, the higher the score. Then, the feature extracted by each selected feature extraction unit 111 is weighted by the score to calculate a weighted average.

[0060] The determining unit 112 may determine, as the data selection feature, each value obtained by merging the feature extracted by each of the selected feature extraction units 111. Specifically, for example, the determining unit 112 determines the data selection feature as follows: Assume that two feature extraction units 111 are selected, and each extracts a 1000-dimensional feature and a 1500-dimensional feature. In this case, the determining unit 112 determines the data selection feature to be a 2500-dimensional feature.

[0061] The determination unit 112 stores the determined feature for data selection in the storage unit 12 in association with the image data from which the feature for data selection was extracted. The determination unit 112 may store the determined feature for data selection in the storage unit 12 in association with the identification number of the image data from which the feature was extracted.

[0062] The determination unit 112 can display the process of determining the features for data selection on the display unit 13. The process of determining the features for data selection includes a feature determination method and various parameters that contributed to determining the features using the determination method. The various parameters that contributed to determining the features for data selection include, for example, the number of clusters of the features extracted by each of the multiple feature extraction units 111 and the number of clusters of the tasks executed by the model. The various parameters that contributed to determining the features for data selection include, for example, the variation in the clusters of the features extracted by each of the multiple feature extraction units 111. The various parameters that contributed to determining the features for data selection include, for example, the distance between the clusters of the features extracted by each of the multiple feature extraction units 111.

[0063] The determination unit 112 can cause the display unit 13 to display a user interface that displays a plurality of feature extraction units 111 in a user-selectable manner. The determination unit 112 can re-determine, as the feature for data selection, the feature extracted by the feature extraction unit 111 selected by the user in the user interface.

[0064] When determining that there is no feature to be selected among the multiple feature extracted by each feature extraction unit 111, the determination unit 112 can output information indicating that there is no feature that can be used to select image data to the display unit 13. This information includes, for example, text data such as "There is no feature that can be used to select image data."

[0065] The data selection unit 200 selects image data that is estimated to be highly effective for learning by the learning unit 300 based on the data selection feature.

[0066] The data selection unit 200 acquires unlabeled image data and labeled image data from the storage unit 12. The data selection unit 200 selects unlabeled image data whose feature amounts satisfy a predetermined standard.

[0067] Fig. 4 is an explanatory diagram for explaining a method for selecting unlabeled image data by the data selection unit 200. In Fig. 4, feature amounts of labeled image data in the feature space are indicated by circles. In Fig. 4, feature amounts of unlabeled image data in the feature space are indicated by crosses.

[0068] The data selection unit 200 calculates the similarity between each of the multiple unlabeled image data to be selected and each of the labeled image data. The data selection unit 200 determines whether the unlabeled image data meets a predetermined standard based on the calculated similarity. The similarity is calculated by comparing the distance between feature vectors, which are the feature amounts of each image data. A known calculation method can be used to calculate the similarity.

[0069] In the example shown in Fig. 4, three labeled image data L1, L2, and L3 are assigned correct labels of class A, class B, and class C, respectively. The data selection unit 200 calculates the distance between each of the two unlabeled image data D1 and D2 and the labeled image data L1 to L3 in the feature space. In the feature space, features that are close to each other have a high similarity, and features that are far from each other have a low similarity.

[0070] Of the three labeled image data L1, L2, and L3, the unlabeled image data D1 is the data that has the highest similarity to the labeled image data L1, which is in class A. The unlabeled image data D2 is the data that has the lowest similarity to any of the three labeled image data L1 to L3, which are in classes A to C, respectively.

[0071] For example, unlabeled image data whose feature values ​​have low similarity to image data of any class is considered to be image data with less characteristics than other unlabeled data. According to this concept, image data whose feature values ​​have high similarity to image data of any class can be said to have a higher learning effect in the learning unit 300. In the example shown in Figure 4, when this concept is adopted, the data selection unit 200 selects unlabeled image data D1.

[0072] On the other hand, unlabeled image data whose feature values ​​have low similarity to image data of any class can be considered to be image data from which the trained model generated by training by the training unit 300 has difficulty determining the correct answer. According to this approach, it can be said that unlabeled image data whose feature values ​​have low similarity to image data of any class has a relatively high learning effect in the training unit 300. On the other hand, it can be said that unlabeled image data whose feature values ​​have high similarity to image data of any class has a relatively low learning effect in the training unit 300. In the example shown in FIG. 4, when this approach is adopted, the data selection unit 200 selects unlabeled image data D2.

[0073] The concept on which the data selection unit 200 selects unlabeled image data can be arbitrarily set by the user at the time of system design.

[0074] The data selection unit 200 stores the distance to each labeled image data and the class of each labeled image data used in selecting the unlabeled image data in the storage unit 12. The data selection unit 200 may store the distance to each labeled image data and the class of each labeled image data in association with the unlabeled image data in the storage unit 12. The data selection unit 200 may also store the distance to each labeled image data and the class of each labeled image data in association with the identification number of the unlabeled image data in the storage unit 12.

[0075] The data selection unit 200 outputs the selected unlabeled image data to an oracle. An oracle is an entity that knows the correct label for the unlabeled image data. The oracle is, for example, a human user of the information processing system 10. The data selection unit 200 displays the selected unlabeled image data on the display unit 13. The data selection unit 200 accepts input of a correct label for the selected unlabeled image data by the user via the input unit 14. The user can view the displayed unlabeled image data and input a correct label for the unlabeled image data.

[0076] The data selection unit 200 associates the selected unlabeled image data with the correct label accepted for the unlabeled image data and stores them in the storage unit 12. The data selection unit 200 may also associate the identification number of the selected unlabeled image data with the correct label accepted for the unlabeled image data and store them in the storage unit 12. In this way, the correct label is assigned to the selected unlabeled image data.

[0077] The selected unlabeled image data can be used as training data by being assigned a correct label. The selected labeled image data can also be used as training data. Note that all labeled image data can be used as training data. A portion of the labeled image data may also be used as training data.

[0078] The data selecting unit 200 outputs the selected unlabeled image data together with the correct label to the learning unit 300 as learning data. The data selecting unit 200 outputs the selected labeled image data together with the correct label to the learning unit 300 as learning data. The data selecting unit 200 can output all labeled image data together with the correct label to the learning unit 300 as learning data. The data selecting unit 200 may also output some labeled image data together with the correct label to the learning unit 300 as learning data.

[0079] The learning unit 300 generates a trained model by learning a model using the learning data output from the data selection unit 200.

[0080] A trained model can be used to perform the task of classifying image data into one of multiple classes. For example, a trained model can input image data of a product and output a classification result of whether the product is good or defective.

[0081] The control unit 11 can add a trained model trained using the image data selected by the data selection unit 200 using the feature determined by the determination unit 112 as a new feature extraction unit 111. The control unit 11 constitutes an extraction unit addition unit.

[0082] 5 is a flowchart showing the operation of the information processing system 10. This flowchart can be executed by the control unit 11 in accordance with a program stored in the storage unit 12.

[0083] The control unit 11 extracts feature amounts from image data using the plurality of feature amount extraction units 111 (S101).

[0084] The control unit 11 selects a feature extraction unit 111 from the plurality of feature extraction units 111 based on the plurality of extracted feature amounts (S102). Specifically, the control unit 11 selects a feature extraction unit 111 that extracts a feature amount for data selection.

[0085] The control unit 11 determines the feature extracted by the selected feature extracting unit 111 as the feature for data selection (S103).

[0086] This embodiment has the following advantages.

[0087] The feature amounts to be used for selecting learning data are determined based on the plurality of feature amounts extracted by the plurality of extraction units, thereby enabling efficient selection of appropriate learning data.

[0088] Furthermore, the extracted features are clustered for each extraction unit, and an extraction unit is selected that extracts features whose cluster count is close to the number of classes of the task executed by the model.The features extracted by the selected extraction unit are then determined as the features to be used for data selection.This allows for the simple and efficient selection of more appropriate learning data.

[0089] Furthermore, the extracted features are clustered for each extraction unit, and the extraction unit that extracts features with minimal variation in each cluster is selected. The features extracted by the selected extraction unit are then determined as the features to be used for data selection. This allows for the simple and efficient selection of more appropriate learning data.

[0090] Furthermore, the system clusters the extracted features for each extraction unit and selects an extraction unit that extracts features with large distances between clusters.The features extracted by the selected extraction unit are then determined as the features to be used for data selection.This allows for the simple and efficient selection of more appropriate learning data.

[0091] Furthermore, an extraction unit that extracts features from labeled data that are similar to features extracted from data with the same label is selected.The feature extracted by the selected extraction unit is then determined as the feature to be used for data selection.This allows for the simple and efficient selection of more appropriate learning data.

[0092] Furthermore, an extraction unit is selected that extracts features from labeled data that are different from features extracted from data with different labels.The features extracted by the selected extraction unit are then determined as the features to be used for data selection.This allows for the simple and efficient selection of more appropriate learning data.

[0093] In addition, the number of extraction units selected is reduced to one, which allows for more flexible and appropriate selection of learning data.

[0094] Furthermore, the feature values ​​extracted by the selected feature extractors are added together to determine the feature values ​​to be used for data selection, thereby enabling more flexible and appropriate selection of learning data.

[0095] Furthermore, the feature values ​​extracted by the selected multiple extraction units are merged and used to select the feature values, which allows for more flexible and appropriate selection of learning data.

[0096] The process of determining the features is also displayed. The extraction units are displayed in a user interface so that they can be selected by the user. The features extracted by the extraction unit selected in the user interface are then redetermined as the features to be used in data selection. This allows for more flexible and appropriate selection of learning data, reflecting the user's intentions.

[0097] The system also determines whether any of the extracted features can be used to select data to be used as model learning data. If it determines that there are no features that can be used to select data to be used as model learning data, it outputs information indicating that there are no features that can be used to select data. This notifies the user that there is no suitable learning data.

[0098] In addition, a model trained using data selected using the determined feature values ​​is added to the extraction unit, making it easier to add candidates for the extraction unit to extract feature values ​​used in data selection.

[0099] The information processing device, control method, and control program according to the present invention are not limited to the above-described embodiments.

[0100] For example, the data may be text data or the like.

[0101] Furthermore, the feature determining unit 100, the data selecting unit 200, and the learning unit 300 may each be configured by separate computers.

[0102] In addition, in the embodiments, part or all of the processing executed by the program may be replaced with hardware such as a circuit.

[0103] While embodiments of the present invention have been described and illustrated in detail, the disclosed embodiments are made for purposes of illustration and example only, and not limitation, and the scope of the present invention should be construed by the language of the appended claims. [Explanation of symbols]

[0104] 10 Information processing systems, 11 control section, 12 storage section, 13 Display section, 14 input section, 15 Communications Department, 100 feature determination unit, 111 feature extraction unit, 112 Decision Section, 200 Data Selection Department, 300 Learning Department.

Claims

1. a plurality of different extraction units each extracting a feature from data; a determination unit that determines, based on a plurality of feature amounts extracted by the plurality of extraction units, feature amounts to be used in selecting the data to be used as learning data for a model; An information processing device having the above.

2. 2. The information processing device according to claim 1, wherein the determination unit clusters the extracted features for each of the extraction units, selects the extraction unit that extracts the features whose number of clusters is close to the number of classes of the task executed by the model, and determines the features extracted by the selected extraction unit as the features to be used in selecting the data.

3. 2. The information processing device according to claim 1, wherein the determination unit clusters the extracted features for each of the extraction units, selects the extraction unit that extracts the features with small variation in each cluster, and determines the features extracted by the selected extraction unit as the features to be used for selecting the data.

4. 2. The information processing device according to claim 1, wherein the determination unit clusters the extracted features for each of the extraction units, selects the extraction unit that extracts the feature having a large distance between clusters, and determines the feature extracted by the selected extraction unit as the feature to be used for selecting the data.

5. 2. The information processing device according to claim 1, wherein the determination unit selects the extraction unit that extracts, from the features extracted from the labeled data, features that are similar to features extracted from the data to which the same label is attached, and determines the features extracted by the selected extraction unit as the features to be used for selecting the data.

6. 2. The information processing device according to claim 1, wherein the determination unit selects the extraction unit that extracts, from the features extracted from the labeled data, features that are different from the features extracted from the data with different labels, and determines the features extracted by the selected extraction unit as the features to be used for selecting the data.

7. The information processing device according to claim 2, wherein the determination unit selects one of the extraction units.

8. The information processing device according to any one of claims 2 to 6, wherein the determination unit determines each value obtained by adding up the feature amounts extracted by the selected plurality of extraction units as the feature amount to be used in selecting the data.

9. The information processing device according to any one of claims 2 to 6, wherein the determination unit determines each value obtained by merging the features extracted by each of the selected plurality of extraction units as the feature to be used for selecting the data.

10. a display unit that displays a process of determining the feature amount by the determination unit; a user interface that displays the extraction unit in a user-selectable manner; The information processing apparatus according to claim 1 , wherein the determination unit redetermines the feature extracted by the extraction unit selected in the user interface as the feature to be used in selecting the data.

11. 2. The information processing device according to claim 1, wherein, when the determination unit determines that there is no feature that can be used to select the data to be used as learning data for the model among the plurality of extracted feature amounts, the determination unit outputs information indicating that there is no feature that can be used to select the data.

12. The information processing apparatus according to claim 1 , further comprising an extraction unit adding unit that adds the model trained using the data selected using the determined feature amount to the extraction unit.

13. (a) extracting features from data using a plurality of different extraction units; (b) determining, based on a plurality of feature amounts extracted by the plurality of extraction units, feature amounts to be used for selecting the data to be used as model learning data; A control method for causing an information processing device to execute the above.

14. 14. The control method according to claim 13, wherein in step (b), the extracted features are clustered for each extraction unit, an extraction unit that extracts the features such that the number of clusters is close to the number of classes of the task executed by the model is selected, and the feature extracted by the selected extraction unit is determined to be the feature to be used for selecting the data.

15. 14. The control method according to claim 13, wherein in step (b), the extracted feature quantities are clustered for each extraction unit, an extraction unit that extracts feature quantities with small variation in each cluster is selected, and the feature quantities extracted by the selected extraction unit are determined to be the feature quantities to be used for selecting the data.

16. 14. The control method according to claim 13, wherein in step (b), the extracted feature quantities are clustered for each extraction unit, an extraction unit that extracts feature quantities with large inter-cluster distances is selected, and the feature quantities extracted by the selected extraction unit are determined to be the feature quantities to be used in selecting the data.

17. 14. The control method according to claim 13, wherein in the step (b), the extraction unit that extracts a feature from the labeled data that is similar to a feature extracted from the data to which the same label is assigned is selected, and the feature extracted by the selected extraction unit is determined to be the feature to be used for selecting the data.

18. 14. The control method according to claim 13, wherein in the step (b), the extraction unit that extracts, from the features extracted from the labeled data, features that are different from the features extracted from the data to which a different label is assigned, is selected, and the feature extracted by the selected extraction unit is determined to be the feature to be used for selecting the data.

19. The control method according to claim 14, wherein in said step (b), one of said extractors is selected.

20. 15. The control method according to claim 14, wherein in the step (b), each value obtained by adding up the feature amounts extracted by the selected plurality of extraction units is determined as the feature amount to be used for selecting the data.

21. 15. The control method according to claim 14, wherein in the step (b), each value obtained by merging the feature amounts extracted by the selected plurality of extraction units is determined as the feature amount to be used for selecting the data.

22. a step (c) of displaying the process of determining the feature amount in the step (b); and (d) displaying the extraction portion via a user interface so that the extraction portion is selectable by a user; 14. The control method according to claim 13, wherein in the step (b), the feature extracted by the extraction unit selected on the user interface is redetermined as the feature to be used in selecting the data.

23. 14. The control method according to claim 13, wherein, in the step (b), when it is determined that there is no feature that can be used to select the data to be used as training data for the model among the plurality of extracted feature amounts, information indicating that there is no feature that can be used to select the data is output.

24. The control method according to claim 13 , further comprising the step (d) of adding to the extractor the model trained using the data selected using the determined feature amount.

25. A control program for causing a computer to execute the control method according to any one of claims 13 to 24.

Images