Information processing device, control method for information processing device, and program
By pre-clustering training data and selecting similar data within clusters, the method addresses biased training data issues, enhancing neural network model accuracy and efficiency.
Patent Information
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- CANON KK
- Filing Date
- 2024-01-22
- Publication Date
- 2026-06-01
AI Technical Summary
Existing machine learning models face challenges in achieving balanced performance across multiple evaluation metrics due to biased training data, leading to inefficient learning and low accuracy in specific data sets.
The method involves pre-clustering training datasets into multiple clusters, identifying problematic data based on differences from ground truth, and selecting additional data within the same cluster with high similarity for training, thereby optimizing the training dataset.
This approach enhances the learning efficiency and accuracy of neural network models by efficiently selecting and refining training data, improving performance on specific evaluation metrics.
Smart Images
Figure 0007868088000002 
Figure 0007868088000003 
Figure 0007868088000004
Abstract
Description
Technical Field
[0001] The present invention particularly relates to an information processing apparatus, a control method for an information processing apparatus, and a program, which are suitable for use in selecting learning data.
Background Art
[0002] In recent years, various services utilizing AI have been provided, and as a method for constructing a model for realizing an AI that predicts an arbitrary event, a method using machine learning is known. As one of the algorithms of a machine learning model, supervised learning using learning data composed of an input and a correct label is known.
[0003] When constructing a model using supervised learning, overfitting can be suppressed and prediction accuracy can be improved by performing learning using high-quality learning data. Here, high-quality learning data represents learning data having a high effect of improving the prediction accuracy of the model. Also, in order to tune a model adapted to a specific situation or application, it is necessary to perform learning using learning data considering the situation or application. Therefore, in supervised learning, it is important to appropriately select the learning data to be used.
[0004] Therefore, a method for excluding unintended data from learning data has been proposed. Patent Document 1 discloses a technique for performing pre-clustering based on feature amounts and class information of object images in image data, and identifying a cluster including incorrect class information by using the average / variance of the distances between a plurality of centroids in the cluster and each feature amount.
Prior Art Documents
Patent Documents
[0005]
Patent Document 1
Non-Patent Documents
[0006] [Non-Patent Document 1] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douse: Deep Clustering for UnsupervisedLearning of Visual Features [Non-Patent Document 2] Krizhevsky, A., Sutskever, I., Hinton, GE: Imagenet classification with deep convolutional neural networks. [Overview of the Initiative] [Problems that the invention aims to solve]
[0007] When a machine learning model evaluates multiple metrics, it may be required to achieve a certain level of performance on any of them. If the performance score on a particular metric is low, it needs to be improved. For example, if the training data is biased, the model may learn efficiently on certain data but not on other data, resulting in a low performance score.
[0008] The method described in Patent Document 1 claims that by excluding unintended data from the training data, training data can be appropriately selected and training can be performed efficiently. However, searching for incorrectly clustered data from all classes is extremely time-consuming. Furthermore, since only information within the cluster can be used, it is not possible to sufficiently improve the training efficiency of the neural network model for specific data.
[0009] In view of the aforementioned problems, the present invention aims to improve the accuracy of neural network models by efficiently selecting training data, improving the non-uniformity of learning efficiency due to the training data, and thereby increasing the accuracy of the neural network model. [Means for solving the problem]
[0010] The information processing device according to the present invention includes a classification means for clustering a training dataset for training a neural network model and classifying it into one of a plurality of clusters, the correct values of the training dataset, and the inference results of the training data by the neural network model or validation data different from that at the time of training. Inference result The system is characterized by comprising: determination means for determining image data whose difference from a threshold is greater than or equal to; calculation means for identifying the cluster to which the image data determined by the determination means belongs from among a plurality of clusters, and calculating the similarity between the data classified into the identified cluster by the classification means and the determined image data; and selection means for selecting data from among the data classified into the identified cluster whose similarity calculated by the calculation means is greater than or equal to a predetermined value as training data for training the neural network model. [Effects of the Invention]
[0011] According to the present invention, it is possible to efficiently select training data, improve the non-uniformity of learning efficiency due to the training data, and improve the accuracy of the neural network model. [Brief explanation of the drawing]
[0012] [Figure 1] This is a block diagram showing an example of the functional configuration of an information processing device according to the embodiment. [Figure 2] This flowchart shows an example of a processing procedure for selecting training data in an embodiment. [Figure 3] This is a diagram illustrating a method for generating a classifier using self-supervised learning. [Figure 4] This is a diagram illustrating the classification method for the training dataset. [Figure 5] This diagram illustrates a method for clustering disliked image data and calculating similarity. [Figure 6] This is a diagram illustrating a method for generating a classifier using supervised learning. [Figure 7] It is a block diagram showing an example of the hardware configuration of an information processing apparatus according to an embodiment.
Mode for Carrying Out the Invention
[0013] Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings. Note that the following embodiments do not limit the invention according to the claims. Although a plurality of features are described in the embodiments, not all of these plurality of features are essential for the invention, and the plurality of features may be arbitrarily combined. Further, in the accompanying drawings, the same or similar configurations are denoted by the same reference numerals, and duplicate descriptions are omitted.
[0014] (First Embodiment) In this embodiment, taking the noise reduction task as an example, the processing flow when training a neural network model (hereinafter referred to as an NN model) will be described. The noise reduction task is a task of estimating a noise-free image (image before degradation) before degradation by noise from a noisy image (degraded image) degraded by noise.
[0015] Here, consider the case of evaluating a plurality of evaluation metrics using an NN model. For example, when evaluating an evaluation metric (PSNR (Peak Signal to Noise Ratio)) regarding less image degradation in the noise reduction task, the PSNR for each of a plurality of regions in the image may be evaluated. In addition, other cases of evaluating a plurality of evaluation metrics include cases such as evaluating the accuracy rate for each class in an object detection task. Thus, when a machine learning model evaluates a plurality of evaluation metrics, a certain or higher evaluation score may be required for any of the plurality of evaluation metrics, and when the evaluation score is low for a specific evaluation metric, it is necessary to improve the corresponding evaluation score.
[0016] As one of the reasons for the low evaluation score in a specific evaluation metric, the difference in the learning efficiency of the NN model can be considered. For example, when there is a bias in the training data, learning progresses efficiently for specific data, and the difference from the correct value (hereinafter referred to as GT (Ground Truth)) is small and the accuracy also increases. On the other hand, for other data, learning does not progress efficiently, and the difference from GT remains large and the accuracy remains low.
[0017] Also, in order to improve the evaluation score in an evaluation metric with a low evaluation score, it is conceivable to replace the training dataset, etc. However, optimizing the training dataset, etc. to improve the evaluation score of a specific evaluation metric takes a lot of time. Therefore, in the present embodiment, the training dataset is clustered in advance and classified into a plurality of clusters, and data is added to the training data according to the similarity from the cluster to which data with a large difference from GT belongs, so that the training data can be selected more efficiently. Hereinafter, the detailed processing in the present embodiment will be described.
[0018] FIG. 7 is a block diagram showing a hardware configuration example of the information processing apparatus 100 according to the present embodiment. In FIG. 7, the processor 701 is, for example, a CPU, and controls the operation of the entire information processing apparatus 100. The memory 702 is, for example, a RAM, and temporarily stores programs, data, etc. The computer-readable storage medium 703 is, for example, a hard disk or a CD-ROM, etc., and stores programs, data, etc. in the long term. In the present embodiment, the program for realizing the functions of each part stored in the storage medium 703 is read into the memory 702. Then, when the processor 701 operates according to the program on the memory 702, the functions of each part are realized.
[0019] In FIG. 7, the input interface 704 is an interface for acquiring information from each device. Also, the output interface 705 is an interface for outputting information to an external device. The bus 706 connects each of the upper parts and enables data exchange.
[0020] Figure 1 is a block diagram showing an example of the functional configuration of the information processing device 100 according to this embodiment. The information processing device 100 includes a model storage unit 101, a training dataset 110, a classifier 120, a dataset group 130, a learning unit 140, a disliked image determination unit 150, a disliked image data 160, a similarity calculation unit 170, and a data selection unit 180.
[0021] The model memory unit 101 stores an NN model for noise reduction. While it is assumed that the NN model is pre-trained, it is not limited to pre-trained models; a pre-trained NN model published by a third party may also be used. Furthermore, the NN model may be a Convolutional Neural Network (CNN) model with convolutional layers, or a Transformer model with an attention mechanism.
[0022] The training dataset 110 is the same dataset used to train the NN model stored in the model memory unit 101. Note that a different image dataset may be used for the training dataset 110. The classifier 120 is a classifier generated to cluster the training dataset 110 and the problematic image data 160, which will be described later. The dataset group 130 represents the result of clustering the training dataset 110 using the classifier 120. The learning unit 140 trains the NN model stored in the model memory unit 101 using the training dataset 110.
[0023] The problematic image determination unit 150 determines problematic image data based on the difference between the results of the training data inferred by the learning unit 140 and the GT of the training dataset 110. Specifically, the problematic image determination unit 150 calculates the difference between the results of the training data inferred by the learning unit 140 and the GT, and if the difference is greater than or equal to a threshold, it defines that training data as problematic image data. The problematic image data 160 is image data that has been determined to be problematic image data by the problematic image determination unit 150.
[0024] The similarity calculation unit 170 determines the cluster to which the problematic image data 160 belongs and calculates the similarity with the data in the dataset group 130 that exists within the same cluster. Based on the similarity calculated by the similarity calculation unit 170, the data selection unit 180 selects additional data from the data in the dataset group 130 within the same cluster to train the NN model.
[0025] Figure 2 is a flowchart showing an example of the process for selecting training data in this embodiment. In step S201, the processor 701 pre-clusters the training dataset 110 using the classifier 120. The clustering method is described in detail below.
[0026] Figure 3 illustrates the method for generating a classifier using self-supervised learning in this embodiment. In this embodiment, the classifier is generated by self-supervised learning called DeepCluster, as described in Non-Patent Literature 1. First, the dataset 301 is input to the classifier 302, and the classifier 302 is trained using pseudo-labels 303 and inference results 304. The learning model used for the classifier 302 is not particularly limited; for example, AlexNet, as described in Non-Patent Literature 2, can be used. Pseudo-labels 303 are generated by inputting the dataset 301 into the classifier 302, obtaining the features before inputting them into the fully connected layer, and then clustering those features using the k-means method. The inference results 304 use the output after inputting the features into the fully connected layer. The loss between the pseudo-labels 303 and the inference results 304 is calculated, and the classifier 302 is trained by backpropagating the result. By using a classifier generation method using self-supervised learning as in this embodiment, it becomes unnecessary to label a large dataset, thus minimizing the burden of preparing training data.
[0027] Figure 4 is a diagram illustrating the classification method of the training dataset 401. Although it is actually a multidimensional feature space, for the sake of simplicity, it will be explained using a two-dimensional feature space. Also, in this embodiment, for the sake of simplicity, the number of clusters will be assumed to be three. Hereafter, the data belonging to clusters 1 to 3 will be referred to as dataset groups 1 to 3, respectively. When the training dataset 401 is input to the classifier 402, the position of each data point in the feature space is identified, and the data is clustered into dataset groups 1 to 3, as shown in the classification result 403.
[0028] Next, in step S202, the learning unit 140 trains the NN model using the training dataset 110. Next, in step S203, the problematic image determination unit 150 determines whether or not problematic image data is included based on the difference between the results of the training data inferred during training in S202 and the GT. As a method for calculating the difference between each inference result and the GT, for example, a loss function such as mean squared error or cross-entropy may be used, or it may be calculated as the difference in pixel values between each inference result and the GT. If all the differences calculated for each inference result are less than the threshold, the problematic image determination unit 150 determines that problematic image data is not included in the inferred training data and terminates the process without taking any further action.
[0029] On the other hand, if there is data where the difference is greater than or equal to a threshold, that data is determined to be a problematic image data. If the problematic image determination unit 150 determines that problematic image data is included in this way, the process proceeds to step S204, and steps S204 to S206 and step S202 are repeated. Note that the range for calculating the difference between each inference result and GT may be the entire image or a local part of the image that has been divided. In this embodiment, the difference with GT was calculated using the inference results from training, but the difference with GT may also be calculated using different data from training, for example, validation data.
[0030] In step S204, the processor 701 inputs the unfavorable image data whose difference from the GT exceeds a threshold into the classifier 120 and performs clustering. Here, the method for clustering the unfavorable image data will be explained.
[0031] Figure 5(a) is a diagram illustrating a clustering method using a classifier. When the disliked image data 501, whose difference from the GT is greater than or equal to a threshold, is input to the classifier 402 and clustering is performed, the classifier 402 outputs the feature vector 502a of the disliked image data 501. Then, it is determined which dataset cluster the output feature vector 502a belongs to in relation to the classification result 403 which was previously clustered. In this embodiment, the feature vector 502a of the disliked image data 501 is assumed to belong to dataset group 3.
[0032] In step S205, the similarity calculation unit 170 calculates the similarity between the problematic image data and each data point in the dataset group belonging to the same cluster. The similarity may be calculated on the feature space obtained when clustering is performed by the classifier, or it may be calculated in an image feature space separate from the classifier, for example, as the difference between the pixel values or brightness values of the images. Alternatively, the image features may be converted into vectors, and the similarity may be calculated using vector distances such as cosine similarity or Euclidean distance. In this embodiment, as an example, cosine similarity is used to calculate the similarity. The cosine similarity cos(x,y) is expressed by the following equation (1).
[0033]
number
[0034] Here, the features of the image data are n-dimensional feature vectors x=(x1,x2,···,x n ) is the feature vector of the datasets belonging to the same cluster identified in step S204, which is an n-dimensional feature vector y=(y1,y2,···y n In this embodiment, the cosine similarity is calculated by substituting the vector values of x and y into equation (1).
[0035] In step S206, the data selection unit 180 selects data with high similarity based on the similarity calculated in step S205 and adds it to the training data. Since cosine similarity approaches 1 as the similarity between two feature vectors increases, data with a cosine similarity value close to 1 is preferentially added to the training data. When adding to the training data, the data may be simply added to the training dataset used for training, or the selected data may be made to have a higher frequency of use during training. In this embodiment, a method for selecting data to add to the training data from the training dataset 110 (dataset group 130) has been described, but training data for addition may be prepared separately, and data may be selected from that.
[0036] Figure 5(b) illustrates an example of selecting data with high similarity based on the similarity calculated in step S205. In this embodiment, first, the similarity between the feature vector 502b of the problematic image data in the image feature space and the data within the dataset group 3, which is in the same cluster, is calculated. Then, data located within a circle 504 with a predetermined threshold 503 as the radius centered on the feature vector 502b of the problematic image data is considered to have a similarity of a predetermined value or higher, and the data located within the circle 504 is selected and added to the training data.
[0037] As in this embodiment, by pre-clustering the training dataset, the calculation of similarity with data belonging to other clusters can be omitted, reducing the time required to select data to add to the training dataset. Furthermore, pre-clustering facilitates data cleansing, such as identifying incorrect clusters using the method described in Patent Document 1, and preventing unintended data from being included in the training dataset, thus enabling the selection of appropriate data as training data.
[0038] In this embodiment, a noise reduction task was used as an example, but it can also be applied to other image quality enhancement tasks such as super-resolution tasks. Furthermore, it is not limited to image quality enhancement tasks, but can also be applied to tasks such as classification tasks and bounding box (BB) detection tasks. In the case of a classification task, training data is selected to improve classes whose accuracy rate is below a predetermined threshold. In the case of a BB detection task, an average precision (AP) is set for each class, and training data is selected to improve classes whose AP is below a predetermined threshold. In addition, this embodiment can be applied to any task that has an evaluation metric that can be expressed numerically.
[0039] (Second embodiment) The following describes the process of selecting training data when training an NN model using a noise reduction task as an example in this embodiment. In the first embodiment, an example was described in which a classifier for clustering a dataset is generated by self-supervised learning as described in Non-Patent Literature 1. In contrast, this embodiment describes an example in which a classifier is generated by supervised learning. The configuration of the information processing device in this embodiment is the same as in Figures 1 and 7, and the basic flow of the processing procedure is the same as in Figure 2, so the explanation is omitted. The difference in this embodiment from the first embodiment is the method of generating the classifier used in step S201.
[0040] Figure 6 illustrates the method for generating a classifier using supervised learning in this embodiment. First, the dataset 601 is input to the classifier 602, and the inference result 603 is output. Then, the loss is calculated between the output inference result 603 and the training data 604, which has been pre-labeled with the correct labels, and the classifier 602 is trained using backpropagation. For the training data 604, for example, data labeled according to the evaluation metric of the noise reduction task is used. Here, in addition to the PSNR mentioned above, other evaluation metrics such as SNR (Signal to Noise Ratio), SSIM (Structual SIMilarity), or MSE (Mean Squared Error) may be used.
[0041] By using the method of this embodiment, if the evaluation score is low for a particular evaluation metric, data similar to the data with the low evaluation score can be immediately identified, allowing for efficient training of the NN model. In this embodiment, the training data 604 was labeled for each evaluation metric, but this is not limited to this. For example, the data may be labeled for each time series to reflect characteristics such as changes over time in the training data, or the data may be labeled for each image characteristic such as brightness, lightness, and saturation to eliminate bias between hues in the training data.
[0042] (Other embodiments) In the embodiments described above, a classifier was generated using either self-supervised learning or supervised learning, but the method of generating the classifier is not limited to these. For example, a classifier may be generated using unsupervised learning to perform clustering according to the similarity of the feature vectors of the images. Furthermore, in the method of generating the classifier, hierarchical clustering such as Ward's method may be used, or representative non-hierarchical clustering such as the k-means method may be used.
[0043] Furthermore, in the embodiment described above, as shown in Figure 5(b), data whose distance from the centroid of the problematic image data is within a predetermined value was selected and added to the training data. However, data to be added to the training data may be selected by a different method. For example, the problematic image data may be clustered to identify the cluster to which the problematic image data belongs, and then data whose distance from the centroid of that cluster is within a predetermined value may be added as training data. By using this method, the calculation of similarity can be made unnecessary, and training data can be selected more quickly.
[0044] The present invention can also be realized by supplying a program that implements one or more of the functions of the above-described embodiments to a system or device via a network or storage medium, and by having one or more processors in the computer of that system or device read and execute the program. It can also be realized by a circuit (e.g., an ASIC) that implements one or more functions.
[0045] This embodiment includes the following configurations, methods, and programs.
[0046] (Composition 1) A classification method that clusters a training dataset for training a neural network model and classifies it into one of several clusters, A determination means for determining image data in which the difference between the correct values of the training dataset and the inference results of the training data by the neural network model or validation data different from the training data is greater than or equal to a threshold, A calculation means that identifies the cluster to which the image data determined by the determination means belongs from among the plurality of clusters, and calculates the similarity between the data classified into the identified cluster by the classification means and the determined image data, A selection means for selecting data from among the data classified into the identified clusters, the data whose similarity calculated by the calculation means is equal to or greater than a predetermined value, as training data for training the neural network model; An information processing device characterized by comprising:
[0047] (Configuration 2) The information processing device according to configuration 1, characterized in that the classification means uses a classifier generated by supervised learning using training data labeled for each evaluation index, each time series, or each image characteristic. (Composition 3) The information processing device according to configuration 2, characterized in that the evaluation index is SNR (Signal to Noise Ratio), PSNR (Peak Signal to Noise Ratio), SSIM (Structural SIMilarity), or MSE (Mean Squared Error). (Composition 4) The information processing apparatus according to configuration 2, characterized in that the aforementioned image characteristics are brightness, lightness, or saturation. (Composition 5) The information processing device according to configuration 1, characterized in that the classification means uses a classifier generated using unsupervised learning to perform classification according to the similarity of the feature vectors of the images. (Composition 6) The information processing apparatus according to configuration 1, characterized in that the classification means classifies using a classifier generated using hierarchical clustering or non-hierarchical clustering.
[0048] (Composition 7) The information processing device according to any one of configurations 1 to 6, characterized in that the determination means divides the learning data into local regions and determines the result by calculating the difference between each local region and the correct answer value. (Composition 8) The information processing apparatus according to any one of configurations 1 to 7, characterized in that the calculation means identifies the cluster to which the determined image data belongs from among the plurality of clusters using the same classifier as the classifier used by the classification means. (Composition 9) The calculation means is an information processing device according to any one of configurations 1 to 8, characterized in that it calculates similarity using image features. (Composition 10) The information processing apparatus according to configuration 9, characterized in that the calculation means converts the feature quantities of the image into vectors and calculates the similarity based on the distance between the vectors. (Composition 11) The information processing apparatus according to any one of configurations 1 to 8, characterized in that the calculation means calculates similarity based on the difference in pixel values or brightness values of the images.
[0049] (Composition 12) A classification method that clusters a training dataset for training a neural network model and classifies it into one of several clusters, A determination means for determining image data in which the difference between the correct values of the training dataset and the inference results of the training data by the neural network model or validation data different from the training data is greater than or equal to a threshold, A means for identifying the cluster to which the image data determined by the determination means belongs from among the plurality of clusters, A selection means for selecting data from among the data classified by the classification means into clusters identified by the identification means, where the distance from the centroid within the identified cluster is within a predetermined value, as training data for training the neural network model. An information processing device characterized by comprising:
[0050] (Method 1) A classification process involves clustering a training dataset for training a neural network model and classifying it into one of several clusters. A determination step for determining image data in which the difference between the correct values of the training dataset and the inference results of the training data by the neural network model or validation data different from the training data is greater than or equal to a threshold, A calculation step which involves identifying the cluster to which the determined image data belongs from among the plurality of clusters in the determination step, and calculating the similarity between the data classified into the identified cluster in the classification step and the determined image data, A selection step in which, from the data classified into the identified clusters, data whose similarity calculated in the calculation step is equal to or greater than a predetermined value is selected as training data for training the neural network model, A control method for an information processing device, characterized by comprising the following:
[0051] (Method 2) A classification process involves clustering a training dataset for training a neural network model and classifying it into one of several clusters. A determination step for determining image data in which the difference between the correct values of the training dataset and the inference results of the training data by the neural network model or validation data different from the training data is greater than or equal to a threshold, A determination step involves identifying the cluster to which the determined image data belongs from among the plurality of clusters, A selection step in which, from the data classified in the classification step into the clusters identified in the specified step, data whose distance from the centroid within the identified cluster is within a predetermined value is selected as training data for training the neural network model; A control method for an information processing device, characterized by comprising the following:
[0052] (Program 1) A classification process involves clustering a training dataset for training a neural network model and classifying it into one of several clusters. A determination step for determining image data in which the difference between the correct values of the training dataset and the inference results of the training data by the neural network model or validation data different from the training data is greater than or equal to a threshold, A calculation step which involves identifying the cluster to which the determined image data belongs from among the plurality of clusters in the determination step, and calculating the similarity between the data classified into the identified cluster in the classification step and the determined image data, A selection step in which, from the data classified into the identified clusters, data whose similarity calculated in the calculation step is equal to or greater than a predetermined value is selected as training data for training the neural network model, A program that causes a computer to execute something.
[0053] (Program 2) A classification process involves clustering a training dataset for training a neural network model and classifying it into one of several clusters. A determination step for determining image data in which the difference between the correct values of the training dataset and the inference results of the training data by the neural network model or validation data different from the training data is greater than or equal to a threshold, A determination step involves identifying the cluster to which the determined image data belongs from among the plurality of clusters, A selection step in which, from the data classified in the classification step into the clusters identified in the specified step, data whose distance from the centroid within the identified cluster is within a predetermined value is selected as training data for training the neural network model; A program that causes a computer to execute something. [Explanation of symbols]
[0054] 120 Classifier, 150 Image detection unit, 170 Similarity calculation unit, 180 Data selection unit
Claims
1. A classification method that clusters a training dataset for training a neural network model and classifies it into one of several clusters, A determination means for determining image data in which the difference between the correct values of the training dataset and the inference results of the training data or the inference results of validation data different from those used during training by the neural network model exceeds a threshold; A calculation means that identifies the cluster to which the image data determined by the determination means belongs from among the plurality of clusters, and calculates the similarity between the data classified into the identified cluster by the classification means and the determined image data. A selection means for selecting data from among the data classified into the identified clusters, the data whose similarity calculated by the calculation means is equal to or greater than a predetermined value, as training data for training the neural network model; An information processing device characterized by comprising:
2. The information processing apparatus according to claim 1, characterized in that the classification means uses a classifier generated by supervised learning using training data labeled for each evaluation index, each time series, or each image characteristic.
3. The information processing apparatus according to claim 2, characterized in that the evaluation index is SNR (Signal to Noise Ratio), PSNR (Peak Signal to Noise Ratio), SSIM (Structural SIMilarity), or MSE (Mean Squared Error).
4. The information processing apparatus according to claim 2, characterized in that the aforementioned image characteristics are brightness, lightness, or saturation.
5. The information processing apparatus according to claim 1, characterized in that the classification means performs classification according to the similarity of the feature vectors of the images using a classifier generated using unsupervised learning.
6. The information processing apparatus according to claim 1, characterized in that the classification means classifies using a classifier generated using hierarchical clustering or non-hierarchical clustering.
7. The information processing apparatus according to claim 1, characterized in that the determination means divides the learning data into local regions and determines the difference between each local region and the correct value.
8. The information processing apparatus according to claim 1, characterized in that the calculation means identifies the cluster to which the determined image data belongs from among the plurality of clusters using the same classifier as the classifier used by the classification means.
9. The information processing apparatus according to claim 1, characterized in that the calculation means calculates similarity using image features.
10. The information processing apparatus according to claim 9, characterized in that the calculation means converts the feature quantities of the image into vectors and calculates the similarity based on the distance between the vectors.
11. The information processing apparatus according to claim 1, characterized in that the calculation means calculates similarity based on the difference in pixel values or brightness values of the images.
12. A classification method that clusters a training dataset for training a neural network model and classifies it into one of several clusters, A determination means for determining image data in which the difference between the correct values of the training dataset and the inference results of the training data or the inference results of validation data different from those used during training by the neural network model exceeds a threshold; A means for identifying the cluster to which the image data determined by the determination means belongs from among the plurality of clusters, A selection means for selecting data from among the data classified by the classification means into clusters identified by the identification means, where the distance from the centroid within the identified cluster is within a predetermined value, as training data for training the neural network model. An information processing device characterized by comprising:
13. A classification process involves clustering a training dataset for training a neural network model and classifying it into one of several clusters. A determination step for determining image data in which the difference between the correct values of the training dataset and the inference results of the training data or the inference results of validation data different from those used during training by the neural network model exceeds a threshold; A calculation step involves identifying the cluster to which the determined image data belongs from among the plurality of clusters, and calculating the similarity between the data classified into the identified cluster and the determined image data in the classification step. A selection step in which, from the data classified into the identified clusters, data whose similarity calculated in the calculation step is equal to or greater than a predetermined value is selected as training data for training the neural network model, A control method for an information processing device, characterized by comprising the following:
14. A classification process involves clustering a training dataset for training a neural network model and classifying it into one of several clusters. A determination step for determining image data in which the difference between the correct values of the training dataset and the inference results of the training data or the inference results of validation data different from those used during training by the neural network model exceeds a threshold; A determination step involves identifying the cluster to which the determined image data belongs from among the plurality of clusters, A selection step in which, from the data classified in the classification step into the clusters identified in the specified step, data whose distance from the centroid within the identified cluster is within a predetermined value is selected as training data for training the neural network model; A control method for an information processing device, characterized by comprising the following:
15. A classification process involves clustering a training dataset for training a neural network model and classifying it into one of several clusters. A determination step for determining image data in which the difference between the correct values of the training dataset and the inference results of the training data or the inference results of validation data different from those used during training by the neural network model exceeds a threshold; A calculation step involves identifying the cluster to which the determined image data belongs from among the plurality of clusters, and calculating the similarity between the data classified into the identified cluster and the determined image data in the classification step. A selection step in which, from the data classified into the identified clusters, data whose similarity calculated in the calculation step is equal to or greater than a predetermined value is selected as training data for training the neural network model, A program that causes a computer to execute something.
16. A classification process involves clustering a training dataset for training a neural network model and classifying it into one of several clusters. A determination step for determining image data in which the difference between the correct values of the training dataset and the inference results of the training data or the inference results of validation data different from those used during training by the neural network model exceeds a threshold; A determination step involves identifying the cluster to which the determined image data belongs from among the plurality of clusters, A selection step in which, from the data classified in the classification step into the clusters identified in the specified step, data whose distance from the centroid within the identified cluster is within a predetermined value is selected as training data for training the neural network model; A program that causes a computer to execute something.