Learning device, learning method, and learning program

Abstract

This learning device comprises: a data extension processing unit that creates learning candidate data increased on the basis of original image data; a certainty factor evaluation unit that attaches a certainty factor evaluation value Es to the learning candidate data; a data for learning extraction unit that extracts data for learning in order of the certainty factor evaluation value Es from the learning candidate data; and a learning processing unit that performs learning based on the data for learning. The data for learning extraction unit extracts the data for learning until the certainty factor evaluation value Es attached to the learning candidate data reaches a preset certainty factor evaluation threshold alpha.

Description

Learning device, learning method, and learning program

[0001] The present invention relates to a learning device, a learning method, and a learning program.

[0002] Conventionally, there has been a deep learning method using a neural network that selects and expands images that are efficient for learning a model to serve as training data from prepared images (for example, Patent Literature 1). The conventional learning method is designed for data expansion processing and data selection processing, and is suitable for learning from large amounts of data.

[0003] In conventional learning methods, the condition that the same image as before the data expansion process should not be selected during the data selection process was added, thereby reducing the overlapping of similar images added by the data expansion process and preventing bias in the training images. However, in conventional learning methods, the number of images used as training data in one learning session is small, making it difficult to fully learn the image features, which increases the number of learning loops and can take a long time to complete learning.

[0004] Japanese Patent Application Publication No. 2024-002431

[0005] An object of the present invention is to provide a learning device, a learning method, and a learning program that enable fast learning.

[0006] (1) A learning device according to one aspect of the present invention includes a data augmentation processing unit that creates expanded learning candidate data based on original image data, a confidence evaluation unit that assigns confidence evaluation values ​​to the learning candidate data, a learning data extraction unit that extracts learning data from the learning candidate data in order of the confidence evaluation values, and a learning processing unit that performs learning based on the learning data, wherein the learning data extraction unit extracts the learning data until the confidence evaluation value assigned to the learning candidate data reaches a predetermined confidence evaluation threshold. (2) In (1) above, the device may include a representative image selection processing unit that selects, from the learning candidate data, a representative image of a group of expanded images that share common original image data. (3) In (2) above, the learning data extraction unit may extract the learning data until the number of types of representative images reaches a predetermined representative image type threshold. (4) In the above (2), the training data extraction unit may include, in the training data, training candidate data assigned a smaller confidence evaluation value than training candidate data assigned the largest confidence evaluation value among the representative images. (5) A training method according to one aspect of the present invention includes: a data augmentation processing step of creating increased training candidate data based on original image data; a confidence evaluation step of assigning confidence evaluation values ​​to the training candidate data; a training data extraction step of extracting training data from the training candidate data in order of the confidence evaluation values; and a training processing step of performing learning based on the training data, wherein in the training data extraction step, the training data is extracted until the confidence evaluation value assigned to the training candidate data becomes equal to a predetermined confidence evaluation threshold.(6) A learning program according to one aspect of the present invention is a learning program that causes a computer to execute a data augmentation processing function that creates increased learning candidate data based on original image data, a confidence evaluation function that assigns a confidence evaluation value to the learning candidate data, a learning data extraction function that extracts learning data from the learning candidate data in order of the confidence evaluation values, and a learning processing function that performs learning based on the learning data, wherein the learning data extraction function extracts the learning data until the confidence evaluation value assigned to the learning candidate data becomes equal to a predetermined confidence evaluation threshold.

[0007] According to the present invention, it is possible to provide a learning device, a learning method, and a learning program that enable fast learning.

[0008] Fig. 1 is a diagram for explaining an overview of a learning device; Fig. 2 is a diagram for explaining an overview of a flow of a learning method; Fig. 3 is a diagram for explaining an image of a learning method when extraction of learning data is limited by the number of representative images; Fig. 4 is a diagram for explaining an image of a learning method when extraction of learning data is limited by a confidence evaluation threshold.

[0009] (Embodiments) Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a diagram illustrating an overview of a learning device 100. FIG. 2 is a diagram illustrating an overview of the flow of a learning method. FIG. 3 is a diagram illustrating an image of a learning method when the extraction of learning data is limited by the number of representative images. FIG. 4 is a diagram illustrating an image of a learning method when the extraction of learning data is limited by a confidence evaluation threshold α. Note that, hereinafter, parts having common functions may be assigned the same reference numerals or symbols.

[0010] (Learning Device) A learning device 100 according to this embodiment will be described with reference to Fig. 1. As shown in Fig. 1, the learning device 100 according to this embodiment includes a data augmentation processing unit 10, a confidence factor evaluation unit 20, a representative image selection processing unit 30, a learning data extraction unit 40, a learning processing unit 50, and a control unit 60.

[0011] The learning device 100 is a device that trains a machine learning model (hereinafter referred to as a model) so that it can appropriately evaluate target data (e.g., image data). Here, a model is a mechanism (algorithm) in machine learning that receives input values ​​that a computer can distinguish and outputs evaluation values.

[0012] The model recognizes and evaluates the content of the target data. For example, if the target data is image data, it evaluates the quality of the subject (whether it is a good product or a defective product, etc.).

[0013] The model has a structure including a neural network and parameters. The neural network has a structure in which multiple neurons are connected. The neural network may be, for example, a hierarchical multi-layer neural network in which layers each include a group of multiple neurons. The neural network is defined by the number of neurons and their connection relationships. The connection strength between neurons or layers is defined using parameters (such as weighting coefficients). Target data is input to the neural network, and the target data is evaluated based on the calculation results of multiple neurons and parameters. The model has, for example, a function to classify target data (training data) into either a good product label or a bad product label, or a function to calculate a good product score.

[0014] The learning device 100 includes a computer (not shown) as its main hardware. The computer includes a processor such as a CPU or GPU, a storage device such as RAM, ROM, a hard disk or solid-state drive (SSD), and a communication unit for connecting to a network such as the Internet. The computer's storage device also stores a learning program for operating the computer as the learning device 100 of this embodiment. Note that the computer also includes a cloud computing system. Each functional unit of the learning device 100 is realized by the computer as hardware and the learning program as software.

[0015] The data augmentation processing unit 10 is a functional unit that generates augmented learning candidate data (learning candidate dataset) based on original image data. This augmentation process improves the quality and quantity of image data, thereby padding the data. This enables the generation of a sufficient amount of high-quality, efficient data for training a machine learning model from a small amount of data, thereby improving the performance of the model. The original image data can be augmented, for example, by a flip process that flips the original image data horizontally, vertically, or rotates it 180 degrees to create data; a shift process that shifts the coordinates of the original image data in the X or Y direction to create data; a micro-rotation process that rotates the original image data by a small angle to the right or left to create data; or a filter process that applies a median filter, Gaussian filter, or other filter to the original image data to create data.

[0016] The confidence evaluation unit 20 is a functional unit that assigns a confidence evaluation value Es, calculated using a predetermined formula, to each piece of candidate learning data. The confidence evaluation value Es is an index that indicates the expected degree to which learning the data will contribute to improving the model's judgment performance. In other words, the confidence evaluation value Es is an index that indicates how confidently the model evaluates (judges) the data at the time the data is learned.

[0017] [Certainty Evaluation Value] When calculating the confidence evaluation value Es, the learning device 100 (certainty evaluation unit 20) calculates, for each piece of candidate learning data, a good-quality distance, which is the distance in feature space between the candidate learning data and the good-quality data, which is training data to which a good-quality label has been assigned (teacher data that served as the basis for machine learning of the model), based on the corresponding feature. In this way, the confidence evaluation value Es may be a parameter calculated based on the distance in feature space between the candidate learning data and the training data. Hereinafter, the "distance" used to express the good-quality distance and the defective distance may be, for example, the Euclidean distance between data projected in feature space. As long as the distance in feature space can be calculated, it is not limited to the Euclidean distance, and Mahalanobis distance or the like may also be used. The distance between a piece of training data, which is one piece of data in the training data group, and a piece of training candidate data, which is one piece of data in the training candidate data group, is calculated, for example, using the following Equation 1: Here, q(k, i) is the coordinate in a certain dimension i of the feature space of training data k, and p(s, i) is the coordinate in a certain dimension i of the feature space of the training candidate data. d(k, s) is the distance between training data k and the training candidate data, the qk vector is a set of coordinate data in the feature space of training data k, and the pk vector is a set of coordinate data in the feature space of the training candidate data. Note that k is an integer equal to or less than the number of data points in the training data (m+n: m and n are integers), i is an integer equal to or less than the predetermined number of dimensions (j) (j is an integer), and s is an integer equal to or less than the number of data points in the training candidate data (t) (t is an integer).

[0018] If the distance between the learning candidate data and the good-quality data OKg, which is one of the good-quality data OK, is d(OKg,s), d(OKg,s) is expressed as the following equation 2 using equation 1. In OKg, OK is a code indicating a good product, and g is an integer equal to or less than the number (m) of good-quality data OK. q(OKg, i) is a coordinate in a certain dimension i of the feature space of the non-defective data OKg in the training data group, and the vector qOKg is a set of coordinate data in the feature space of the non-defective data OKg.

[0019] If the set of distances between the learning candidate data and each of the non-defective data OK is a vector d(OK, s), the vector d(OK, s) is expressed as the following equation 3 using equation 2.

[0020] The non-defective distance E(OK, s) in the learning candidate data is, for example, the smallest value among the elements of the vector d(OK, s). That is, the non-defective distance E(OK, s) is the smallest value among the elements of the vector d(OK, s), which is a set of distances between the learning candidate data and each non-defective data OK. The non-defective distance E(OK, s) is expressed as the following equation 4 using equation 3. In this case, the smaller the non-defective distance E(OK, s), the closer the learning candidate data is to one of the non-defective data OK in the feature space.

[0021] The non-defective distance E(OK, s) in the learning candidate data may be calculated by, for example, extracting a smallest elements from the elements of the vector d(OK, s) and averaging the a elements. a is a natural number, for example, 3. In this case, the non-defective distance E(OK, s) is expressed as the following equation 5 using equation 3. In this case, the smaller the non-defective distance E(OK, s), the closer the learning candidate data is to a plurality (a pieces) of non-defective data in the feature space, and the closer the learning candidate data is to a group of non-defective data (non-defective cluster).

[0022] Furthermore, the learning device 100 (certainty evaluation unit 20) calculates, for each piece of learning candidate data, a defective product distance, which is the distance in feature space between the learning candidate data and defective product data, which is learning data to which a defective product label has been assigned, based on the corresponding feature. If the distance between the learning candidate data and the defective product data NGh among the defective product data NG is d(NGh, s), d(NGh, s) is expressed as the following equation 6 using equation 1. In NGh, NG is a code indicating a defective product, and h is an integer equal to or less than the number (n) of data in the defective product data NG. Note that q(NGh, i) is a coordinate in a certain dimension i of the feature space of the defective product data NGh among the learning data, and the vector qNGh is a set of coordinate data in the feature space of the defective product data NGh.

[0023] If the set of distances between the learning candidate data and each defective data NG is a vector d(NG, s), the vector d(NG, s) is expressed as the following equation 7 using equation 6.

[0024] The defective product distance E(NG, s) in the learning candidate data is, for example, the smallest value among the elements of the vector d(NG, s). That is, the defective product distance E(NG, s) is the smallest value among the distances between the learning candidate data and each defective product data NG. The defective product distance E(NG, s) is expressed as the following equation 8 using equation 7. In this case, the smaller the defective product distance E(NG, s), the closer the learning candidate data is located to one of the defective product data NG in the feature space.

[0025] The defective product distance E(NG, s) in the learning candidate data may be calculated by, for example, extracting a smallest elements from the elements of the vector d(NG, s) and averaging these a elements. In this case, the defective product distance E(NG, s) is expressed as the following equation 9 using equation 7. In this case, the smaller the defective product distance E(NG, s), the closer the learning candidate data is to a plurality (a pieces) of NG defective product data in the feature space, and the closer the learning candidate data is to a group of NG defective product data (defective product cluster).

[0026] Furthermore, the learning device 100 (certainty evaluation unit 20) calculates a certainty evaluation value Es for the learning candidate data using the calculated non-defective distance E(OK, s) and defective distance E(NG, s). The certainty evaluation value Es is, for example, the value obtained by dividing the non-defective distance E(OK, s) by the defective distance E(NG, s), and is expressed as in the following Equation 10:

[0027] For example, the smaller the confidence evaluation value Es is than 1, the smaller the good-product distance E(OK, s) is compared to the defective-product distance E(NG, s), indicating that the learning candidate data is closer to the good-product cluster than to the defective-product cluster. Therefore, when the learning candidate data is data with a defective product label, the smaller the confidence evaluation value Es is, the more difficult it is for the learning candidate data to be classified into a good product label or a defective product label in the model based on the learning results of the current training data group, indicating that the data has a high learning effect for the model.

[0028] On the other hand, for example, the greater the confidence evaluation value Es is, the smaller the defective distance E(NG, s) is compared to the good distance E(OK, s), indicating that the learning candidate data is closer to the defective cluster than to the good cluster. Therefore, when the learning candidate data is data with a good label, the greater the confidence evaluation value Es is, the more difficult it is for the learning candidate data to be classified into a good label or a bad label in the model based on the learning results of the current training data group, indicating that the data has a high learning effect for the model.

[0029] The representative image selection processor 30 is a functional unit that selects, from the training candidate data, a representative image representing a group of extended images that share common original image data. For example, as shown in Figures 3 and 4, three circular marks in the training candidate data constitute one extended image group, and the one with the smallest certainty evaluation value Es (when sorted in ascending order) among the three circular marks is the representative image of the extended image group. When the training candidate data are sorted in ascending order of certainty evaluation value Es, the representative image selection processor 30 selects the first circular shape, the third triangular shape, the sixth diamond shape, and the nth chestnut shape in the order of the training candidate data.

[0030] The learning data extraction unit 40 is a functional unit that performs extraction processing to calculate a confidence evaluation value Es for a learning candidate data set as a population and extract, as learning data, learning candidate data whose confidence evaluation value Es is within a predetermined range (for example, equal to or less than a predetermined confidence evaluation threshold α). In other words, the learning data extraction unit 40 extracts learning data from the learning candidate data in order of the confidence evaluation value Es.

[0031] The learning processing unit 50 is a functional unit that causes the model to undergo learning processing at a predetermined learning rate. The learning processing is processing that adjusts the parameters of the model so that they approach optimal values.

[0032] The control unit 60 is a functional unit that controls the learning data extraction unit 40 and the learning processing unit 50 to perform effective learning processing for the model.

[0033] (Learning Method) Next, the flow of the learning method according to the embodiment will be described. The learning method can be implemented using the learning device 100. Each step (process or process) in the learning method can be executed by a corresponding unit (each module) in the learning device 100. FIG. 2 is a diagram illustrating an overview of the flow of the learning method. FIG. 3 is a diagram illustrating an image of the learning method when the extraction of learning data is limited by the number of representative images. FIG. 4 is a diagram illustrating an image of the learning method when the extraction of learning data is limited by a confidence evaluation threshold α.

[0034] (1) In advance, the learning device 100 acquires a dataset that will serve as the original image data.

[0035] (2) Next, as shown in Figure 2, the learning device 100 uses the data augmentation processing unit 10 to create expanded learning candidate data (learning candidate dataset) based on the original image data (data augmentation processing step) (Process A). This process can improve the quality and quantity of the image data, thereby padding it. This allows for the generation of a large amount of data that is efficient and high-quality for learning a machine learning model from a small amount of data, thereby improving the performance of the model.

[0036] (3) Next, the learning device 100 assigns a confidence evaluation value Es calculated by a predetermined formula to each piece of training candidate data (confidence evaluation step) using the confidence evaluation value Es assigned to each piece of training candidate data (image) using the representative image selection processing unit 30. The learning device 100 then selects a representative image from the training candidate data set as a representative image from a group of extended images with common original image data using the confidence evaluation value Es assigned to each piece of training candidate data (image) using the representative image selection processing unit 30 (process B). Here, the representative image may be the image with the smallest confidence evaluation value Es among multiple pieces of image data (extended image group) extended from the common original image data (i.e., an image for which no images with the same extended source data were selected in process B). In other words, process B selects the image with the smallest evaluation value when images are sorted in ascending order of evaluation value as the representative image, and does not select subsequent images with the same original image data as representative images. For example, as shown in FIG. 3 , the round image in the first row of the “training candidate data” column is selected as the representative image, while the round images in the second and fourth rows, which have the same original image data, are not selected as representative images.

[0037] (3-1) Here, the learning data extraction unit 40 may extract learning data until the certainty evaluation value Es assigned to the candidate learning data becomes equal to a predetermined certainty evaluation threshold value α (learning data extraction step).

[0038] (4) Next, the number of representative images selected in process B is detected. If a representative image is detected, the process proceeds to process D. If a representative image is not detected, the process ends learning (process C).

[0039] (5) Here, the selected representative images are sorted in order of the confidence evaluation value Es (e.g., ascending order), and for example, based on a predetermined value, images with smaller confidence evaluation values ​​Es are used for learning and images with larger confidence evaluation values ​​Es are used for other purposes, and the images are classified as either for learning or other purposes (Process D).

[0040] (5-1) Here, in process D, the learning device 100 may select representative images until the certainty evaluation value Es reaches a preset certainty evaluation threshold value α (Condition 1). This makes it difficult for inefficient candidate learning data to be extracted as learning data (teacher data), and enables the extraction of an appropriate amount of learning data that is neither too little nor too much.

[0041] (5-2) Alternatively, in process D, the learning device 100 may perform the classification process until the number of representative images reaches a predetermined specified number (condition 2). In other words, the learning data extraction unit 40 of the learning device 100 continues to extract learning data until the number of types of representative images reaches a predetermined threshold value β for the number of types of representative images. This allows an appropriate amount of learning data to be extracted, neither too little nor too much.

[0042] (5-3) Alternatively, in process D, the learning device 100 may execute the process until all representative images are classified (condition 3). This allows an appropriate amount of learning data to be extracted, neither too little nor too much. Note that the above conditions 1 to 3 may be applied independently or in any combination.

[0043] (6) The representative images classified for learning by the process D are extracted as learning data (process E).

[0044] (6-1) Here, images with a smaller confidence evaluation value Es than the representative image are extracted as training data (process G). In other words, the training data extraction unit 40 includes in the training data training candidate data assigned a smaller confidence evaluation value Es than the training candidate data assigned the largest confidence evaluation value Es among the representative images. This makes it possible to extract all of the extended images that share original image data with the representative image, thereby reducing the shortage of training data and extracting a sufficient amount of training data with a high learning effect, compared to when other training candidate data that share original image data with the representative image is not extracted.

[0045] (7) After process E, learning is performed based on the learning data extracted up to process E (process F). Note that a termination condition is set to stop learning when a predetermined number of times is exceeded, which is expected to result in proper learning. After learning in process F, the process returns to process B.

[0046] (8) As appropriate, between process A and process B, a pre-trained model is set as an initial value (process H).

[0047] In this way, the learning method according to this embodiment uses the learning device 100 to extract learning data until it satisfies predetermined conditions (conditions 1 to 3 described above) based on a representative image in the original image data from which data is expanded. This ensures that the high-quality learning data necessary for learning is extracted in a sufficient amount that is not excessively large, while still extracting different types of high-quality learning data necessary for learning. This makes it possible to provide a learning device 100 and learning method that can learn quickly.

[0048] (Learning Program) The learning program has the function of causing a computer to execute each process performed at each step in the above-described learning method. Specifically, the learning program has a data augmentation processing function, a confidence assessment function, a learning data extraction function, and a learning processing function, corresponding to each step in the above-described learning method. This makes it possible to generate high-quality, efficient data and a sufficient large amount of data for learning a machine learning model from a small amount of data, thereby improving the performance of the model. Therefore, a learning program with fast learning can be provided.

[0049] The technical scope of the present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the present invention.

[0050] In addition, the components in the above-described embodiments may be replaced with known components as appropriate without departing from the spirit of the present invention. Furthermore, the above-described modifications may be combined as appropriate without departing from the spirit of the present invention.

[0051] As described above, the learning device 100 according to the embodiment includes a data augmentation processing unit 10 that creates expanded learning candidate data based on original image data, a confidence evaluation unit 20 that assigns a confidence evaluation value Es to the learning candidate data, a learning data extraction unit 40 that extracts learning data from the learning candidate data in order of the confidence evaluation value Es, and a learning processing unit 50 that performs learning based on the learning data. The learning data extraction unit 40 extracts learning data until the confidence evaluation value Es assigned to the learning candidate data reaches a predetermined confidence evaluation threshold α. This makes it possible to generate a large amount of high-quality, efficient data for learning a machine learning model from a small amount of data, thereby improving the performance of the model. Therefore, a learning device 100 with fast learning speed can be provided.

[0052] A learning method according to an embodiment includes a data expansion processing step of creating expanded learning candidate data based on original image data, a confidence evaluation step of assigning a confidence evaluation value Es to the learning candidate data, a learning data extraction step of extracting learning data from the learning candidate data in order of the confidence evaluation value Es, and a learning processing step of performing learning based on the learning data. In the learning data extraction step, the learning method extracts learning data until the confidence evaluation value Es assigned to the learning candidate data reaches a predetermined confidence evaluation threshold α. This makes it possible to generate a large amount of data that is efficient and high-quality for learning a machine learning model from a small amount of data, thereby improving the performance of the model. This provides a learning method that enables fast learning.

[0053] A learning program according to an embodiment causes a computer to execute a data augmentation processing function that creates expanded learning candidate data based on original image data, a confidence evaluation function that assigns a confidence evaluation value Es to the learning candidate data, a learning data extraction function that extracts learning data from the learning candidate data in order of the confidence evaluation value Es, and a learning processing function that performs learning based on the learning data. In the learning data extraction function, the learning program extracts learning data until the confidence evaluation value Es assigned to the learning candidate data reaches a predetermined confidence evaluation threshold α. This enables the generation of a large amount of high-quality, efficient data for learning a machine learning model from a small amount of data, thereby improving the performance of the model. This provides a learning program that enables fast learning.

[0054] 100 Learning device, 10 Data expansion processing unit, 20 Confidence evaluation unit, 30 Representative image selection processing unit, 40 Learning data extraction unit, 50 Learning processing unit, 60 Control unit, α Confidence evaluation threshold, β Representative image type number threshold

Claims

1. A learning device comprising: a data augmentation processing unit that creates increased learning candidate data based on original image data; a confidence evaluation unit that assigns a confidence evaluation value to the learning candidate data; a learning data extraction unit that extracts learning data from the learning candidate data in order of the confidence evaluation value; a learning processing unit that performs learning based on the learning data; and a representative image selection processing unit that selects, from the learning candidate data, a representative image of a group of extended images that share the original image data and extracts it as the learning data, wherein the learning data extraction unit extracts the learning data until the confidence evaluation value assigned to the learning candidate data reaches a predetermined confidence evaluation threshold.

2. The learning device according to claim 1, wherein the learning data extraction unit extracts the learning data until the number of types of representative images reaches a preset threshold value for the number of types of representative images.

3. The learning device according to claim 1, wherein the learning data extraction unit includes in the learning data learning candidate data that has a smaller certainty evaluation value than the learning candidate data that has the largest certainty evaluation value among the representative images.

4. A learning method comprising: a data augmentation processing step of creating increased learning candidate data based on original image data; a confidence evaluation step of assigning confidence evaluation values ​​to the learning candidate data; a learning data extraction step of extracting learning data from the learning candidate data in order of the confidence evaluation values; a learning processing step of performing learning based on the learning data; and a representative image selection processing step of selecting, from the learning candidate data, a representative image of a group of extended images that share the original image data, and extracting it as the learning data, wherein in the learning data extraction step, the learning data is extracted until the confidence evaluation value assigned to the learning candidate data reaches a predetermined confidence evaluation threshold.

5. A learning program that causes a computer to execute: a data augmentation processing function that creates increased learning candidate data based on original image data; a confidence evaluation function that assigns a confidence evaluation value to the learning candidate data; a learning data extraction function that extracts learning data from the learning candidate data in order of the confidence evaluation value; a learning processing function that performs learning based on the learning data; and a representative image selection processing function that selects, from the learning candidate data, a representative image of a group of extended images that share the original image data, and extracts it as the learning data, wherein the learning data extraction function extracts the learning data until the confidence evaluation value assigned to the learning candidate data reaches a predetermined confidence evaluation threshold.

Images