JP7695097B2 - Sampling device and sampling method - Google Patents

Description

The present invention relates to a technique for sampling data to be labeled from a large amount of data.

Image classification and object detection using machine learning such as deep learning require a huge number of images, anywhere from thousands to millions. These images must then be labeled to identify the type of object, the object detection area, and so on. Even if the data is not images, labeling is still necessary for data classification and data recognition using machine learning.

JP 2020-119238 A

To reduce the labor required for this labeling work, it is possible to use learning methods such as semi-supervised learning, which uses both labeled and unlabeled data for training. However, with this learning method, if there is a bias in the data to be labeled, sampling bias will occur. If sampling bias occurs in the labeled data, bias will also occur in learning using unlabeled data, resulting in a decrease in the accuracy of the trained model for image recognition.

One idea for reducing the occurrence of sampling bias is to increase the amount of data to be labeled. However, if the amount of data to be labeled is increased, the labor required for labeling will not be significantly reduced, and the possibility of bias in the labeled data will still remain.

The invention described in Patent Document 1 selects more uncertain data as the target for labeling, and does not suppress the occurrence of the sampling bias described above.

In view of the above problems, the present invention aims to provide a sampling technique that can suppress the occurrence of sampling bias while suppressing the number of samples.

The sampling device according to the present invention is configured (first configuration) to include a clustering unit that acquires a feature vector for each of a plurality of data and classifies the plurality of data into a plurality of clusters based on the distance of the feature vector, and a selection unit that distinguishes, for each of the plurality of clusters, a first data group that is close to the center of gravity of the cluster and a second data group other than the first data group, and selects the data from the first data group.

In the sampling device of the first configuration, the selection unit may be configured to select the data closest to the center of gravity of the cluster (second configuration).

In the sampling device of the first or second configuration, the selection unit may be configured to divide the second data group into a third data group that is far from the center of gravity of the cluster and a fourth data group other than the third data group, and select the data from the third data group (third configuration).

In the sampling device of the third configuration described above, the selection section may be configured to select, from the third data group, data that is farther away from the center of gravity of the cluster with higher priority (fourth configuration).

In the sampling device of any one of the first to fourth configurations described above, the selection unit may be configured (fifth configuration) to select a larger number of data from each of the plurality of clusters as the variance of the feature vectors within the cluster increases.

In the sampling device of any one of the first to fourth configurations described above, the selection unit may be configured (sixth configuration) to select a larger number of data from each of the plurality of clusters as the number of data in the cluster increases.

The sampling method according to the present invention is configured (seventh configuration) to include a clustering step of acquiring a feature vector for each of a plurality of data and classifying the plurality of data into a plurality of clusters based on the distance of the feature vector, and a selection step of distinguishing, for each of the plurality of clusters, a first data group close to the center of gravity of the cluster and a second data group other than the first data group, and selecting the data from the first data group.

The present invention makes it possible to reduce the number of samples while suppressing the occurrence of sampling bias.

FIG. 1 is a diagram showing an example of a schematic configuration of an information processing apparatus according to an embodiment; A flowchart showing an example of the operation of the information processing device according to the embodiment. FIG. 1 is a diagram showing an example of a result of the clustering unit classifying a plurality of data into first to fourth clusters; FIG. 1 is a diagram showing a schematic diagram of a result of discriminating a first data group close to the center of gravity of a first cluster from a second data group other than the first data group; FIG. 1 is a diagram showing a schematic diagram of a result of dividing a second data group into a third data group far from the center of gravity of the first cluster and a fourth data group other than the third data group;

An exemplary embodiment of the present invention will be described in detail below with reference to the drawings.

1. Configuration of information processing device
1 is a diagram showing a schematic configuration example of an information processing device according to an embodiment. The information processing device 1 is an example of a sampling device. The information processing device 1 may be an information processing device installed in a single location, or may be a distributed information processing device in which components are distributed and installed in multiple locations.

The information processing device 1 includes a control unit 11 and a memory unit 12.

The control unit 11 is a computer having at least one processor. Specifically, the control unit 11 is a computer having a CPU (Central Processing Unit), a RAM (Random Access Memory), and a ROM (Read Only Memory), which are not shown. The control unit 11 processes and transmits/receives information based on the programs stored in the storage unit 12, and controls the entire information processing device 1.

The control unit 11 includes a clustering unit 11a and a selection unit 11b. The CPU executes calculations according to the programs stored in the storage unit 12, thereby realizing various functions of the control unit 11, such as the clustering unit 11a.

The clustering unit 11a obtains a feature vector for each of the multiple data sets and classifies the multiple data sets into multiple clusters based on the distance between the feature vectors.

The selection unit 11b distinguishes between a first data group that is close to the center of gravity of each of the multiple clusters and a second data group other than the first data group, and selects data from the first data group.

2. Operation of the information processing device
2 is a flowchart showing an example of the operation of the information processing device 1. When the information processing device 1 is turned on and a plurality of data are input, the information processing device 1 starts the operation of the flowchart shown in FIG. 2. The input form of the plurality of data is not particularly limited. The information processing device 1 may receive the plurality of data by wireless communication, may receive the plurality of data by wired communication, or may input the plurality of data via a removable storage medium to the information processing device 1. The plurality of data input to the information processing device 1 is composed of, for example, several thousand to several million images. Note that the plurality of data input to the information processing device 1 may be data other than images.

First, the clustering unit 11a acquires a feature vector for each of the multiple data (step S10). The feature vector is acquired by, for example, machine learning. This machine learning may be supervised machine learning or unsupervised machine learning. Furthermore, the number of dimensions of the feature vector is not particularly limited. In the following explanation, in order to simplify the explanation and the figures, a case where the feature vector is a two-dimensional vector is given as an example. As an example of unsupervised machine learning, a DNN (Deep Neural Network) or the like can be used.

Next, the clustering unit 11a classifies the multiple data into multiple clusters based on the distance of the feature vectors (step S20). The clustering unit 11a performs, for example, hierarchical clustering, and classifies the multiple data into multiple clusters by fixing the number of clusters to be finally classified. As the distance of the feature vectors, for example, Euclidean distance, cosine similarity, etc. can be used.

Figure 3 is a diagram showing an example of the results of the clustering unit 11a classifying multiple data into first to fourth clusters CL1 to CL4. Each black circle in Figure 3 represents an individual piece of data, but to simplify the explanation, only a smaller number of data than the actual number is shown. The explanation will continue below using the example shown in Figure 3.

Next, for each of the multiple clusters, the selection unit 11b distinguishes between a first data group close to the center of gravity of the cluster and a second data group other than the first data group, and selects the data from the first data group (step S30). This allows typical data for each cluster to be selected, making it possible to suppress the occurrence of sampling bias while suppressing the number of samples. Note that the center of gravity of a cluster means the center of gravity of the feature vectors of the data belonging to the cluster.

FIG. 4 is a diagram showing a schematic diagram of a result of distinguishing a first data group G1 close to the center of gravity C1 of the first cluster CL1 from a second data group G2 other than the first data group G1. In FIG. 4, the first data group G1 includes three pieces of data, but the method of determining the boundary between the first data group G1 and the second data group G2 is not particularly limited. The number of pieces of data belonging to the first data group G1 may be determined, for example, taking into consideration the number of clusters, the number of pieces of data belonging to the clusters, and the number of pieces of data selected by the selection unit 11b. When the number of pieces of data belonging to the first data group G1 is determined based on the number of pieces of data belonging to the clusters, for example, the number of pieces of data in the first data group G1 is 20% of the number of pieces of data belonging to the clusters, and the number of pieces of data in the second data group G2 is 80% of the number of pieces of data belonging to the clusters.

The selection unit 11b may also select data belonging to the first data group G1 as data within a first predetermined distance from the center of gravity C1 of the cluster. The value of the first predetermined distance may also be determined taking into consideration the number of clusters and the number of data selected by the selection unit 11b. By determining the value in this manner, it is possible to select typical data for a cluster and suppress sampling bias caused by selecting multiple data.

In this embodiment, the selector 11b selects the data closest to the center of gravity of the cluster. In other words, the most representative data of each cluster is selected, so that the occurrence of sampling bias can be further suppressed while suppressing the number of samples.

To explain this in more detail using the example shown in Figure 3, in this embodiment, the selection unit 11b selects, with top priority, the data D1 (see Figure 4) from the first cluster CL1 that is closest to the center of gravity C1 of the first cluster CL1, selects, with top priority, the data from the second cluster CL2 that is closest to the center of gravity C2 of the second cluster CL2, selects, with top priority, the data from the third cluster CL3 that is closest to the center of gravity C3 of the third cluster CL3, and selects, with top priority, the data from the fourth cluster CL4 that is closest to the center of gravity C4 of the fourth cluster CL4.

Refer to Figures 2 and 5. Furthermore, the selection unit 11b divides the second data group G2 into a third data group G3 that is far from the center of gravity of the cluster, and a fourth data group G4 that is close to the center of gravity of the cluster. The selection unit 11b also selects data from the third data group (step S30). The division into the third data group G3 and the fourth data group G4 may be performed based on the distance from the center of gravity of the cluster or the number of data. This allows data near the boundary of each cluster (the part farthest from the center of gravity of each cluster) to be selected, making it possible to efficiently reduce the number of samples while suppressing the occurrence of sampling bias.

Using FIG. 5, the division of the third data group G3 and the fourth data group G4 will be described. FIG. 5 is a diagram showing a schematic diagram of the result of dividing the second data group G2 into the third data group G3 far from the center of gravity C1 of the first cluster CL1 and the fourth data group G4 other than the third data group G3, with respect to the first cluster CL1. The method of determining the boundary between the third data group G3 and the fourth data group G4 is not particularly limited. For example, the number of data belonging to the third data group G3 may be determined taking into consideration the number of clusters, the number of data belonging to the second data group G2, and the number of data selected by the selection unit 11b. When the number of data belonging to the third data group G3 is determined based on the number of data belonging to the second data group G2, for example, the number of data belonging to the third data group G3 is 50% of the number of data belonging to the second data group G2. Also, for example, data belonging to the third data group G3 may be defined as data that is at a second predetermined distance or more from the center of gravity of the cluster, and the value of the second predetermined distance may be determined taking into account the number of clusters and the number of data selected by the selection unit 11b.

In this embodiment, the selector 11b preferentially selects data from the third data group that is farther away from the center of gravity of the cluster. In other words, among the special (non-typical) data of each cluster, the more special the data is, the more preferentially it is selected, so that it is possible to further suppress the occurrence of sampling bias while suppressing the number of samples. Note that, unlike this embodiment, for example, the selector 11b may preferentially select data from the third data group that is closer to the center of gravity of the cluster, or, for example, the selector 11b may randomly select a predetermined number of data from the third data group.

The data selection method will be specifically described using the examples shown in Figures 3 and 5. In this embodiment, the selector 11b selects at least one data from the first data group G1 of the first cluster CL1 (a data group in the area closest to the center of gravity C1 of the first cluster CL1). Examples of the method of selecting data from the first data group G1 include (1) a method of giving priority to data closer to the center of gravity C1 of the first cluster CL1, (2) a method of giving priority to data farther from the center of gravity C1 of the first cluster CL1, and (3) a method of randomly selecting data from the first data group G1. Similarly, data is selected from the data group in the area closest to the center of gravity of the cluster for the second to fourth clusters CL2 to CL4.

In this embodiment, when data selection from the data group in the area closest to the center of gravity of the cluster is completed, the selector 11b selects at least one data from the third data group G3 of the first cluster CL1 (the data group in the area farthest from the center of gravity C1 of the first cluster CL1). Examples of methods for selecting data from the third data group G3 include (1) a method of giving priority to data that is farther from the center of gravity C1 of the first cluster CL1, (2) a method of giving priority to data that is closer to the center of gravity C1 of the first cluster CL1, and (3) a method of randomly selecting data from the third data group G3. Similarly, data is selected from the data group in the area farthest from the center of gravity of the cluster for the second to fourth clusters CL2 to CL4.

In this embodiment, if the number of data selected by the selection unit 11b does not reach a preset value even after data selection from the data group in the area closest to the center of gravity of the cluster and the data group 3 in the area farthest from the center of gravity of the cluster is completed, the selection unit 11b selects at least one data from the fourth data group G4 of the first cluster CL1 (the data group in the area second farthest from the center of gravity C1 of the first cluster CL1). Examples of methods for selecting data from the fourth data group G4 include (1) a method of giving priority to data that is farther from the center of gravity C1 of the first cluster CL1, (2) a method of giving priority to data that is closer to the center of gravity C1 of the first cluster CL1, and (3) a method of randomly selecting data from the fourth data group G4. Similarly, data is selected from the data group in the area second farthest from the center of gravity of the cluster for the second to fourth clusters CL2 to CL4.

For example, the selector 11b selects first the data from the area closest to the center of gravity of the cluster, second the data from the area farthest from the center of gravity of the cluster, and third the data from the area second farthest from the center of gravity of the cluster. By selecting in this manner, it is possible to further reduce the number of samples and suppress sampling bias.

When the number of data selected by the selection unit 11b reaches a preset value, the selection unit 11b ends the selection process. For example, if the number of data selected by the selection unit 11b is set to 100 and multiple data are classified into the first to fourth clusters CL1 to CL4 as in the example shown in FIG. 3, the selection unit 11b selects 25 data from each of the first to fourth clusters CL1 to CL4.

When the processing of step S30 ends and the data selected by the selection unit 11b is output, the information processing device 1 ends the operation of the flowchart shown in FIG. 2. The output form of the data selected by the selection unit 11b is not particularly limited. The information processing device 1 may transmit the data selected by the selection unit 11b via wireless communication, may transmit the data selected by the selection unit 11b via wired communication, or may output the data selected by the selection unit 11b via a storage medium that is detachable from the information processing device 1.

3. Modifications
The above-described embodiments should be considered in all respects as illustrative and not restrictive, and the technical scope of the present invention is indicated by the claims, not by the description of the above-described embodiments, and should be understood to include all modifications that fall within the meaning and scope of the claims.

For example, the selection unit 11b may select a larger number of data from each of a plurality of clusters as the variance of the feature vectors within the cluster increases. This is expected to efficiently suppress the occurrence of sampling bias in each cluster. For example, according to the variances V1 to V4 of the feature vectors within the first to fourth clusters CL1 to CL4, the selection unit 11b may select data such that the number of data selected from the first cluster CL1: the number of data selected from the second cluster CL2: the number of data selected from the third cluster CL3: the number of data selected from the fourth cluster CL4 = V1:V2:V3:V4.

For example, the selection unit 11b may select a larger number of data from each of the multiple clusters as the number of data in the cluster increases. This is expected to efficiently suppress the occurrence of sampling bias in each cluster. For example, depending on the numbers of data N1 to N4 in the first to fourth clusters CL1 to CL4, the selection unit 11b may select data such that the number of data selected from the first cluster CL1: the number of data selected from the second cluster CL2: the number of data selected from the third cluster CL3: the number of data selected from the fourth cluster CL4 = N1:N2:N3:N4.

In the above embodiment, an example was described in which the control unit 11 and the storage unit 12 are separate configurations, but the present invention is not limited to this. The storage unit 12 may be included in the control unit 11. Also, the storage unit 12 may be a storage medium that can be attached to and detached from the control unit 11. Here, the recording medium is, for example, a flexible disk, a hard disk, a CD-ROM, an MO, a DVD, a DVD-ROM, a DVD-RAM, a large-capacity DVD, a next-generation DVD, or a semiconductor memory.

Reference Signs List 1 Information processing device 11 Control unit 11a Clustering unit 11b Selection unit 12 Storage unit

Claims (6)

a clustering unit that acquires a feature vector for each of a plurality of data and classifies the plurality of data into a plurality of clusters based on a distance between the feature vectors;
a selection unit that distinguishes, for each of the plurality of clusters, a first data group close to a center of gravity of the cluster and a second data group other than the first data group, selects the data from the first data group, and increases the number of the data selected from each of the plurality of clusters as the variance of the feature vectors within the cluster increases ;
A sampling device comprising: The sampling device according to claim 1, wherein the selection unit selects the data closest to the center of gravity of the cluster. The sampling device according to claim 1 or 2, wherein the selection unit divides the second data group into a third data group that is far from the center of gravity of the cluster and a fourth data group that is close to the center of gravity of the cluster, and selects the data from the third data group. The sampling device according to claim 3 , wherein the selection section selects, from the third data group, the data that is farther away from the center of gravity of the cluster with a higher priority. The sampling device according to any one of claims 1 to 4, wherein the selection unit increases the number of data selected from each of the plurality of clusters as the number of data in the cluster increases. a clustering step of acquiring a feature vector for each of a plurality of data, and classifying the plurality of data into a plurality of clusters based on the distance between the feature vectors;
a selection step of distinguishing, for each of the plurality of clusters, a first data group close to a center of gravity of the cluster and a second data group other than the first data group, and selecting the data from the first data group, and increasing the number of the data selected from each of the plurality of clusters as the variance of the feature vectors within the cluster increases ;
The sampling method comprises: