JPWO2020054551A1 - Information processing equipment, information processing methods, programs - Google Patents

Abstract

The information processing apparatus 100 of the present invention includes a feature amount extraction unit 110 that extracts a first feature amount of training data using a neural network model, and information corresponding to a class to which the training data set in the neural network model belongs. Based on this, the clustering unit 120 that converts the first feature amount into the second feature amount and clusters the second feature amount, and the target for modifying the class based on the clustering result of the second feature amount. It is provided with a correction target selection unit 130 for selecting the learning data.

Description

The present invention relates to an information processing device, an information processing method, and a program that support correction of labeling of learning data used in a neural net.

In recent years, machine learning using neural networks has been used in various fields. In order to create an inference model using a neural network, a large amount of training data is required, but due to the large amount of training data, labeling is performed by multiple people or in the long term. For this reason, the training data may be labeled with ambiguous judgment criteria. Such labeling reduces the inference accuracy of the inference model.

Label modification is effective as a method for improving the accuracy of such an inference model. For this reason, an expert visually confirms the learning data and corrects the label while arranging the criteria for labeling.

International Publication No. 2017/179258

However, since there are a large amount of learning data to be checked for labels and they are not organized, there arises a problem that it takes man-hours to correct the labels. As a technique related to such a problem, there is a technique of Patent Document 1. In Patent Document 1, the man-hours required for label correction are reduced by changing the image display method according to the difference between the clustering of the likelihood vectors and the average in each cluster.

Here, the feature quantity obtained from the neural network has a property that those with different labels are separated from each other in the direction perpendicular to the identification plane. Therefore, as shown in FIG. 1, there is a problem that similar data with different labels C1 and C2, that is, data labeled with ambiguous criteria C1 and C2 are difficult to be grouped in the same cluster. .. However, a solution to such a problem is not described in Patent Document 1. As a result, there is still a problem that it is not possible to improve the labeling accuracy while reducing the man-hours required for label correction of the training data.

Therefore, an object of the present invention is information that can solve the problem that in machine learning using a neural network, it is not possible to reduce the number of steps required for label correction of training data and improve the labeling accuracy. The purpose is to provide processing devices, information processing methods, and programs.

The information processing device, which is one embodiment of the present invention, is
A feature extraction unit that extracts the first feature of the training data using a neural network model,
A clustering unit that converts the first feature amount into a second feature amount and clusters the second feature amount based on the information corresponding to the class to which the learning data is set in the neural network model. When,
A correction target selection unit that selects the learning data to be modified based on the clustering result of the second feature amount, and a modification target selection unit.
With,
It takes the configuration.

Further, the program which is one form of the present invention is
For information processing equipment
A feature extraction unit that extracts the first feature of the training data using a neural network model,
A clustering unit that converts the first feature amount into a second feature amount and clusters the second feature amount based on the information corresponding to the class to which the learning data is set in the neural network model. When,
A correction target selection unit that selects the learning data to be modified based on the clustering result of the second feature amount, and a modification target selection unit.
To realize,
It takes the configuration.

Further, the information processing method, which is one form of the present invention, is
Extract the first feature of the training data using the neural network model,
Based on the information set in the neural network model and corresponding to the class to which the learning data belongs, the first feature amount is converted into the second feature amount, and the second feature amount is clustered.
The training data to be modified in the class is selected based on the clustering result of the second feature amount.
It takes the configuration.

By configuring as described above, the present invention can reduce the man-hours required for label correction of learning data and improve the labeling accuracy in machine learning using a neural net.

It is a figure for demonstrating the problem in machine learning. It is a block diagram which shows the structure of the label correction support device in Embodiment 1 of this invention. It is a figure for demonstrating the state of the process by the label correction target presenting method determination apparatus disclosed in FIG. It is a figure for demonstrating the state of the processing by the presentation / correction apparatus disclosed in FIG. It is a flowchart which shows the operation of the label correction support device disclosed in FIG. It is a flowchart which shows the operation of the label correction support device disclosed in FIG. It is a block diagram which shows the structure of the label correction support device in Embodiment 2 of this invention. It is a block diagram which shows the structure of the label correction support device in Embodiment 3 of this invention. It is a block diagram which shows the structure of the information processing apparatus in Embodiment 4 of this invention.

<Embodiment 1>
The first embodiment of the present invention will be described with reference to FIGS. 2 to 6. FIG. 2 is a diagram for explaining the configuration of the label correction support device, and FIGS. 3 to 6 are diagrams for explaining the operation of the label correction support device.

[composition]
The present invention includes a label correction support device 1 as shown in FIG. The label correction support device 1 is composed of one or a plurality of information processing devices including an arithmetic unit and a storage device 10. The label correction support device 1 includes a label correction target presentation method determining device 20 and a presentation / correction device 30 constructed by the arithmetic unit executing a program. The label correction target presentation method determination device 20 includes a feature amount extraction device 21, an image sorting device 22, a clustering device 23, and a cluster sorting / sorting device 24. Hereinafter, each configuration will be described in detail.

First, the label correction support device 1 of the present invention supports correction of labeling of learning data used in a neural network. Specifically, the label correction support device 1 has a function of presenting learning data to the user, obtaining feedback, and correcting the label of the learning data, as will be described later. Therefore, the label correction support device 1 can be applied to a multi-class classification problem in which the classes to which the learning data belong are mutually exclusive. This problem is, for example, a problem of classifying whether the image is a "dog" or a "cat" when the target learning data is an image. Hereinafter, the number of classes to be classified is C (C = 2 in the above example). However, the learning data targeted by the present invention is not limited to images.

The storage device 10 stores learning data, a neural network model, and various set values. At this time, the learning data is composed of a set of an image and a label, and there are a plurality of learning data.

The label correction target presentation method determination device 20 uses the learning data from the storage device 10, the neural network model, and various setting values as inputs, and as will be described later, the determination criteria for labeling some similar image clusters are ambiguous. It is a device that outputs items arranged in order from the item.

The feature amount extraction device 21 (feature amount extraction unit) receives the learning data from the storage device 10 and the neural network model as inputs, and the feature amount x_n (n = 1, ..., N) of each of the N images as the training data. ) (First feature amount). The feature amount extraction device 21 extracts an input vector to the classification layer in front of the classification activation layer located at the end of the neural network as a feature amount. At this time, the classification layer is a hyperplane w_c · x + b = 0 (c = 1, ..., C) with C classes, with the feature vector x_n having the output of each neuron in the immediately preceding layer as an input. ) Is a layer that outputs a C-dimensional classification vector z that has each element as an element. The classification activation layer is a layer that takes the classification vector z as an input and outputs a C-dimensional vector y in which the value of each dimension is activated by an activation function such as a Softmax function. At this time, each element y_c (c = 1, ... C) of this vector is a certainty y_c indicating how much the input image seems to be a class corresponding to that element (certainty for the class to which the learning data belongs). Can be regarded as.

The image selection device 22 (feature amount selection unit) inputs the feature amount x_n from the feature amount extraction device 21 and various setting values from the storage device 10, and the feature amount of the image labeled with an ambiguous determination standard. It is a device that selects and outputs only x_m (m = 1, ..., M). Here, M is the number of features after selection. This selection can be realized by an arbitrary method using the above-mentioned certainty y_c and feature amount x_n. For example, a method of selecting whether or not the certainty y_c is included in a preset range, or a method of setting the range as a value calculated from the average value or variance of the certainty y_c can be considered. As an example, a feature amount whose certainty y_c is lower than the threshold value as a set value stored in the storage device 10 is selected. As a result, as will be described later, the image sorting device 22 clusters the images labeled with the clear judgment criteria generated in the clustering device 23 into the same cluster as the images labeled with the ambiguous judgment criteria. It plays a role in suppressing the problem. As the setting value required for sorting, the image sorting setting value in various setting values is used.

The clustering device 23 (clustering unit) receives the neural network model from the storage device 10 and various set values, and the selected feature amount x_m from the image sorting device 22 as inputs, and obtains training data based on the neural network model. Cluster the feature x_m. Then, the clustering result (each feature amount, cluster ID for each class) k_m, c (m = 1, ..., M: c = 1, ..., C) is output. Specifically, the clustering apparatus 23 first transforms the selected features x_m by orthographically projecting them onto the identification plane corresponding to each class c of the neural network model. That is, as shown in FIG. 3, by orthographically projecting the feature amount of the learning data onto the identification plane corresponding to the class to which the learning data belongs, the dimension of the feature amount is compressed and converted. Orthographic projection onto the identification plane is performed by, for example, the formula shown in FIG. Next, the clustering device 23 clusters the converted features x'_m, c (second feature) using the neural network model. At this time, as a clustering method, for example, a method using a generally known k-means method, Mean-Shift method, or the like can be considered.

As described above, the clustering device 23 has a function of converting the selected feature amount x_m, and such a function is such that the above-mentioned "feature amounts obtained from the neural network are those with different labels." Since it has the property of being vertically separated from the identification plane, it is difficult for similar images with different labels to be grouped together in the same cluster, that is, images labeled with ambiguous criteria. Is playing. At this time, the information in the certainty direction is lost due to the orthogonal projection onto the identification plane of the above-mentioned class, but the above-mentioned image sorting device 22 has already selected the feature amount to be converted, and plays a role of supplementing such information. There is. As the setting value required by the k-means method or the like, the clustering setting value in various setting values is used.

The cluster sorting / sorting device 24 (correction target selection unit) has various set values from the storage device 10, selected feature amounts x_m from the image sorting device 22, and clustering results k_m, c from the clustering device 23. Is a device that outputs cluster sorting information by inputting. Specifically, the clustering sorting / sorting device 24 assigns the randomness of the label given to the feature amount in the clustered cluster to the cluster k as the ambiguity a_k, and uses the ambiguity a_k and the class c as a reference. Arrange the cluster k and exclude the cluster k that does not meet the presentation conditions. That is, the learning data belonging to the cluster that is not excluded is selected as the target for modifying the class to which the learning data belongs. At this time, as a method of arranging, a method of arranging in the order of class c and then arranging in order from the one having the highest ambiguity a_k, or a method of ignoring the class and arranging in order from the one having the largest ambiguity a_k can be considered.

As a method of calculating the ambiguity a_k by the cluster sorting / sorting device 24, for example, a method using entropy for labeling in the cluster, a method using a simple ratio of labeling, and the like can be considered. Further, as a method of calculating the presentation condition, for example, a method by comparing the ambiguity a_k and the threshold value, a method by comparing the rank of the ambiguity a_k of the cluster k and the threshold value, and the like can be considered. In addition, as a method of setting the threshold value of the ambiguity a_k, for example, a method of simply using a fixed value or a method of adding a fixed value in the ambiguity a_k'of another cluster can be considered. As the setting value required by this device, the cluster selection / sorting setting value in various setting values is used.

The presentation / correction device 30 (class setting unit) inputs the learning data from the storage device 10, the clustering result from the clustering device 23, and the cluster sorting information from the cluster sorting / sorting device 24, and is a correction target for the user. The image to be modified, which is the training data of the above, and the label indicating the class to which the image belongs are presented in order for each cluster. Then, the presentation / correction device 30 updates the label representing the class to which the learning data stored in the storage device 10 belongs, based on the correction information of the label input from the user for the presented correction target image. Set. Here, an example of the presentation / correction method is shown in FIG. By the method described above, the training data and the label, which are similar images whose labeling criteria are ambiguous, are presented together. In this example, the user is asked to enter the modified label in the "?" Mark field. It also has an interface that displays the previous or next cluster by pressing the "arrow" mark.

[motion]
Next, the operation of the label correction support device 1 described above will be described with reference to the flowcharts of FIGS. 5 to 6. Note that FIG. 5 shows the overall operation of the label correction support device 1, and FIG. 6 shows the operation of the clustering device 23.

First, the feature amount extraction device 21 takes the training data from the storage device 10 and the neural network model as inputs, and extracts the feature amount x_n (n = 1, ..., N) (first feature amount) of the training data (first feature amount). Step S1). At this time, the feature amount extraction device 21 also calculates the certainty y_c for the class to which the learning data belongs based on the feature amount of the learning data.

Subsequently, the image sorting device 22 sorts only the feature amount x_m (m = 1, ..., M) of the image that can be determined to be labeled by an ambiguous judgment criterion by using the feature amount x_n and the certainty y_c ( Step S2). Then, the clustering apparatus 23 performs the feature amount conversion (step S3) and clustering (step S4) described below using the selected feature amount.

Here, the processing operation by the clustering apparatus 23 will be described with reference to the flowchart of FIG. First, the clustering device 23 receives the neural network model and the selected feature amount x_m (m = 1, ..., M) from the storage device 10 and the image selection device 22, respectively (step S11). At this time, M is the number of features selected by the image sorting device 22. Further, the neural network model includes a weight parameter w_c and a bias parameter b_c corresponding to each class c (= 1, ..., C) of the classification layer.

Subsequently, the clustering device 23 initializes the variable, that is, c = 1 (step S12). Subsequently, the clustering device 23 orthographically projects the selected feature amount onto the identification plane (step S13). Specifically, using the weight parameter w_c and the bias parameter b_c of the neural network model, all the selected features x_m are positive to the corresponding class c identification plane using the equations shown in FIG. Convert to projected features x'_m, c (second feature).

Subsequently, the clustering apparatus 23 clusters the orthographically projected features x'_m, c by the k-means method or the Mean-Shift method (step S14). As a result, the cluster ID k_m, c (= 1, ..., K) to which the orthographically projected features x'_m, c belong can be obtained. At this time, K is the number of clusters.

After that, the clustering apparatus 23 updates the value of c to c + 1 (step S15), and if c <= C (No in step S16), clusters the next class (steps S13 to S15). If c = C (Yes in step S16), the clustering result (k_m, c (m = 1, ..., M: c = 1, ... C)) is sent to the cluster sorting / sorting device 24 (step S17). ..

Subsequently, the cluster sorting / sorting device 24 assigns the randomness of the label given to the feature amount in the clustered cluster to the cluster k as the ambiguity a_k, and clusters based on the ambiguity a_k and the class c. Arrange k and select an image that is the training data for which the class is to be modified (step S5).

Then, as shown in FIG. 4, the presentation / correction device 30 presents the image and the label, which are the learning data to be corrected, to the user in order for each cluster (step S6). The presentation / correction device 30 updates and sets a label representing a class to which the learning data stored in the storage device 10 belongs, based on the correction information of the label input from the user for the presented correction target image. (Step S7).

As described above, in the present invention, the features obtained from the neural net have the property that those with different labels are separated from each other in the direction perpendicular to the identification plane. That is, it is possible to solve the problem that the images labeled by an ambiguous criterion are difficult to be grouped in the same cluster. The reason is that the present invention has a function of clustering the orthogonal projections of the selected features on the identification plane and determining the method of presenting the correction target by selecting and rearranging the clusters.

<Embodiment 2>
Next, a second embodiment of the present invention will be described with reference to FIG. FIG. 7 is a diagram showing a configuration of the label correction support device 1 in the present embodiment.

The label correction support device 1 in the present embodiment includes an automatic label correction device 40 constructed by the arithmetic unit executing a program. The label automatic correction device 40 is a device that updates the learning data of the storage device 10 by inputting the clustering result from the clustering device 23 and the cluster sorting information from the cluster sorting / sorting device 24. At this time, as an update method, for example, a method of randomly selecting a label for each presented cluster and updating all the labels of the images in the cluster to the selected label, or all the images in the cluster. A method of selecting a label and updating the label depending on whether or not the average value of the certainty of the above exceeds the threshold value can be considered. As described above, in the present embodiment, the label of the image is updated all at once and automatically, so that the label can be easily corrected.

<Embodiment 3>
Next, a third embodiment of the present invention will be described with reference to FIG. FIG. 8 is a diagram showing the configuration of the label correction support device 1 in the present embodiment.

The label correction support device 1 in the present embodiment further includes a set value update device 50 constructed by the arithmetic unit executing a program, in addition to the configuration provided in the label correction support device 1 described in the first embodiment. The setting value updating device 50 has a function of presenting various setting values to the user, receiving the updated values from the user, and updating the various setting values stored in the storage device 10 based on the updated values as described above. Have.

<Embodiment 4>
Next, a fourth embodiment of the present invention will be described with reference to FIG. FIG. 9 is a block diagram showing the configuration of the information processing apparatus according to the fourth embodiment. In this embodiment, the outline of the configuration of the label correction support device 1 described in the first embodiment is shown.

As shown in FIG. 9, the information processing device 100 in this embodiment is
The feature amount extraction unit 110 that extracts the first feature amount of the training data using the neural network model, and the feature amount extraction unit 110.
A clustering unit that converts the first feature amount into a second feature amount and clusters the second feature amount based on the information corresponding to the class to which the learning data is set in the neural network model. 120 and
A correction target selection unit 130 that selects the learning data to be corrected based on the clustering result of the second feature amount, and a correction target selection unit 130.
To be equipped.

The feature amount extraction unit 110, the clustering unit 120, and the correction target selection unit 130 are realized by the information processing apparatus executing a program.

Then, the information processing device 100 having the above configuration is
Extract the first feature of the training data using the neural network model,
Based on the information set in the neural network model and corresponding to the class to which the learning data belongs, the first feature amount is converted into the second feature amount, and the second feature amount is clustered.
The training data to be modified in the class is selected based on the clustering result of the second feature amount.
It operates to execute the process.

According to the above invention, the first feature amount of the learning data is converted into the second feature amount based on the information corresponding to the class to which the learning data belongs and clustered, so that the learning to be the target of modifying the class is performed. You can select the data. As a result, it is possible to reduce the man-hours required for label correction of learning data and improve the labeling accuracy.

<Additional notes>
Part or all of the above embodiments may also be described as in the appendix below. Hereinafter, the outline of the configuration of the information processing device, the information processing method, and the program in the present invention will be described. However, the present invention is not limited to the following configurations.

(Appendix 1)
A feature extraction unit that extracts the first feature of the training data using a neural network model,
A clustering unit that converts the first feature amount into a second feature amount and clusters the second feature amount based on the information corresponding to the class to which the learning data is set in the neural network model. When,
A correction target selection unit that selects the learning data to be modified based on the clustering result of the second feature amount, and a modification target selection unit.
Information processing device equipped with.

(Appendix 2)
The information processing device according to Appendix 1.
The clustering unit compresses the dimension of the first feature amount and converts it into the second feature amount.
Information processing device.

(Appendix 3)
The information processing device according to Appendix 1 or 2.
The clustering unit orthographically projects the first feature amount onto the identification plane corresponding to the class to which the learning data belongs in the neural network model, and converts the first feature amount into the second feature amount. ,
Information processing device.

(Appendix 4)
The information processing device according to any one of Supplementary note 1 to 3.
A feature amount selection unit for selecting the first feature amount based on the value based on the first feature amount is further provided.
The clustering unit converts the selected first feature amount into the second feature amount.
Information processing device.

(Appendix 5)
The information processing device according to Appendix 4.
The feature amount selection unit selects the first feature amount based on a value representing the certainty for the class to which the learning data belongs based on the first feature amount.
Information processing device.

(Appendix 6)
The information processing device according to any one of Appendix 1 to 5.
The correction target selection unit is a target for modifying a class based on the randomness of the class to which the learning data, which is the source of the second feature amount, belongs in the cluster to which the second feature amount is clustered. Select the training data,
Information processing device.

(Appendix 7)
The information processing device according to any one of Supplementary note 1 to 6.
A class setting unit for changing and setting the class to which the learning data selected as the target for modifying the class belongs is provided.
Information processing device.

(Appendix 8)
For information processing equipment
A feature extraction unit that extracts the first feature of the training data using a neural network model,
A clustering unit that converts the first feature amount into a second feature amount and clusters the second feature amount based on the information corresponding to the class to which the learning data is set in the neural network model. When,
A correction target selection unit that selects the learning data to be modified based on the clustering result of the second feature amount, and a modification target selection unit.
A program to realize.

(Appendix 8.1)
The program described in Appendix 8
In the information processing device
Further, a feature amount selection unit that selects the first feature amount based on the value based on the first feature amount is further realized.
The clustering unit converts the selected first feature amount into the second feature amount.
program. ..

(Appendix 8.2)
To the program described in Appendix 8 or 8.1,
A program for further realizing a class setting unit for changing and setting a class to which the learning data selected as a target for modifying a class belongs to the information processing device.

(Appendix 9)
Extract the first feature of the training data using the neural network model,
Based on the information set in the neural network model and corresponding to the class to which the learning data belongs, the first feature amount is converted into the second feature amount, and the second feature amount is clustered.
The training data to be modified in the class is selected based on the clustering result of the second feature amount.
Information processing method.

(Appendix 10)
The information processing method described in Appendix 9
The dimension of the first feature is compressed and converted into the second feature, and the second feature is clustered.
Information processing method.

(Appendix 11)
The information processing method according to Appendix 9 or 10.
The first feature amount is orthographically projected onto the identification plane corresponding to the class to which the learning data belongs in the neural network model, the first feature amount is converted into the second feature amount, and the second feature amount is converted into the second feature amount. Clustering features,
Information processing method.

(Appendix 12)
The information processing method according to any one of Supplementary note 9 to 11.
The first feature amount is selected based on the value based on the first feature amount, and the first feature amount is selected.
The selected first feature amount is converted into the second feature amount, and the second feature amount is clustered.
Information processing method.

(Appendix 13)
The information processing method described in Appendix 12
The first feature amount is selected based on a value representing the certainty for the class to which the learning data belongs based on the first feature amount.
Information processing method.

(Appendix 14)
The information processing method according to any one of Supplementary note 9 to 13.
Based on the randomness of the class to which the learning data, which is the source of the second feature, belongs in the cluster to which the clustered second feature belongs, the learning data to be modified is selected. ,
Information processing method.

(Appendix 15)
The information processing method according to any one of Supplementary note 9 to 14.
Change and set the class to which the learning data selected as the target for modifying the class belongs.
Information processing method.

The above-mentioned program can be stored and supplied to a computer using various types of non-transitory computer readable medium. Non-temporary computer-readable media include various types of tangible storage media. Examples of non-temporary computer-readable media include magnetic recording media (eg, flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (eg, magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs. CD-R / W, semiconductor memory (for example, mask ROM, PROM (Programmable ROM), EPROM (Erasable PROM), flash ROM, RAM (Random Access Memory)) are included. The program may also be supplied to the computer by various types of transient computer readable medium. Examples of temporary computer-readable media include electrical, optical, and electromagnetic waves. The temporary computer-readable medium can supply the program to the computer via a wired communication path such as an electric wire and an optical fiber, or a wireless communication path.

Although the invention of the present application has been described above with reference to the above-described embodiment and the like, the present invention is not limited to the above-described embodiment. Various changes that can be understood by those skilled in the art can be made within the scope of the invention of the present application in terms of the configuration and details of the invention of the present application.

The present invention enjoys the benefit of priority claim based on the patent application of Japanese Patent Application No. 2018-169829, which was filed in Japan on September 11, 2018, and is described in the patent application. All contents are included in this specification.

1 Label correction support device 10 Storage device 20 Label correction target presentation method determination device 21 Feature extraction device 22 Image sorting device 23 Clustering device 24 Cluster sorting / sorting device 30 Presentation / correction device 40 Label automatic correction device 50 Setting value update device 100 Information processing device 110 Feature extraction unit 120 Clustering unit 130 Correction target selection unit

Claims (15)

A feature extraction unit that extracts the first feature of the training data using a neural network model,
A clustering unit that converts the first feature amount into a second feature amount and clusters the second feature amount based on the information corresponding to the class to which the learning data is set in the neural network model. When,
A correction target selection unit that selects the learning data to be modified based on the clustering result of the second feature amount, and a modification target selection unit.
Information processing device equipped with. The information processing device according to claim 1.
The clustering unit compresses the dimension of the first feature amount and converts it into the second feature amount.
Information processing device. The information processing device according to claim 1 or 2.
The clustering unit orthographically projects the first feature amount onto the identification plane corresponding to the class to which the learning data belongs in the neural network model, and converts the first feature amount into the second feature amount. ,
Information processing device. The information processing device according to any one of claims 1 to 3.
A feature amount selection unit for selecting the first feature amount based on the value based on the first feature amount is further provided.
The clustering unit converts the selected first feature amount into the second feature amount.
Information processing device. The information processing device according to claim 4.
The feature amount selection unit selects the first feature amount based on a value representing the certainty for the class to which the learning data belongs based on the first feature amount.
Information processing device. The information processing device according to any one of claims 1 to 5.
The correction target selection unit is a target for modifying a class based on the randomness of the class to which the learning data, which is the source of the second feature amount, belongs in the cluster to which the second feature amount is clustered. Select the training data,
Information processing device. The information processing device according to any one of claims 1 to 6.
A class setting unit for changing and setting the class to which the learning data selected as the target for modifying the class belongs is provided.
Information processing device. For information processing equipment
A feature extraction unit that extracts the first feature of the training data using a neural network model,
A clustering unit that converts the first feature amount into a second feature amount and clusters the second feature amount based on the information corresponding to the class to which the learning data is set in the neural network model. When,
A correction target selection unit that selects the learning data to be modified based on the clustering result of the second feature amount, and a modification target selection unit.
A program to realize. Extract the first feature of the training data using the neural network model,
Based on the information set in the neural network model and corresponding to the class to which the learning data belongs, the first feature amount is converted into the second feature amount, and the second feature amount is clustered.
The training data to be modified in the class is selected based on the clustering result of the second feature amount.
Information processing method. The information processing method according to claim 9.
The dimension of the first feature is compressed and converted into the second feature, and the second feature is clustered.
Information processing method. The information processing method according to claim 9 or 10.
The first feature amount is orthographically projected onto the identification plane corresponding to the class to which the learning data belongs in the neural network model, the first feature amount is converted into the second feature amount, and the second feature amount is converted into the second feature amount. Clustering features,
Information processing method. The information processing method according to any one of claims 9 to 11.
The first feature amount is selected based on the value based on the first feature amount, and the first feature amount is selected.
The selected first feature amount is converted into the second feature amount, and the second feature amount is clustered.
Information processing method. The information processing method according to claim 12.
The first feature amount is selected based on a value representing the certainty for the class to which the learning data belongs based on the first feature amount.
Information processing method. The information processing method according to any one of claims 9 to 13.
Based on the randomness of the class to which the learning data, which is the source of the second feature, belongs in the cluster to which the clustered second feature belongs, the learning data to be modified is selected. ,
Information processing method. The information processing method according to any one of claims 9 to 14.
Change and set the class to which the learning data selected as the target for modifying the class belongs.
Information processing method.

Images