Learning device, learning method, and learning program product

Abstract

[Subject] Provide a learning device, a learning method, and a learning program for appropriately learning a model. [Solution] The learning device 1 is a device for learning to make appropriate judgments in a machine learning model M that calculates an evaluation value for target data and judges it as a good product (first label) or a defective product (second label). The learning device 1 includes a data acquisition unit 2, an addition processing unit 3, a control unit 4, an extraction unit 5, a learning unit 6, a display unit 7, and a label change unit 8. The data to be processed are teaching data 11 assigned a first label or a second label, teaching candidate data 12 that becomes a candidate for the teaching data 11, and changed teaching candidate data 13 that is subjected to addition processing by the addition processing unit 3. The data that becomes the matrix 14 is acquired by the data acquisition unit 2, the data is added by the addition processing unit 3, and the extraction unit 5 extracts the extracted data 15. Before the extraction processing is performed by the extraction unit 5, learning can be performed by adding data by the addition processing unit 3.

Description

Technical Field

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

[0002] Conventionally, image processing techniques have been used to utilize the learning results obtained through deep learning using neural networks to determine the quality of products. Furthermore, to improve the efficiency of learning performed by deep learning, Patent Document 1 proposes a learning support device and a learning device for appropriately learning a model.

[0003] The learning support device of Patent Document 1 is based on a model that uses and learns teacher data to classify target data into either a first label or a second label, and teacher data having first data assigned a first label and second data assigned a second label, and deriving feature quantities of the teacher data for each teacher data.

[0004] In addition, the same device is provided with: an export unit that exports the feature value of the professor candidate data for each professor candidate data based on the model and at least one professor candidate data respectively assigned with either the first label or the second label; a calculation unit that calculates at least one of the distance between the professor candidate data and the first data and the distance between the professor candidate data and the second data for each professor candidate data; and a selection unit that selects data to be added as professor data from the professor candidate data based on the distance.

[0005] In the learning support device of patent document 1, professor candidate data whose distance is closest to professor data of the second label is selected from a plurality of professor candidate data classified as the first label, and professor candidate data whose distance is closest to professor data of the first label is selected from a plurality of professor candidate data classified as the second label.

[0006] This process allows selection of the professor candidate data closest to the second label from among the professor candidate data with the first label. In other words, it allows selection of professor candidate data that is difficult to identify as either the first label or the second label. The same applies to professor candidate data with the second label. In the learning support device of Patent Document 1, by converting such difficult-to-identify professor candidate data into professor data, the quality of the professor data can be improved, enabling efficient model learning.

[0007] Prior Art Literature

[0008] Patent Literature

[0009] Patent Document 1: Japanese Patent Application Publication No. 2021-103344 Summary of the Invention

[0010] Problems to be solved by the invention

[0011] The learning support device described in Patent Document 1 enables efficient learning. Incidentally, in general learning methods, it is known that increasing the amount of teaching data using a specified technique can improve learning outcomes. Therefore, in the learning support device described in Patent Document 1, it is also expected that increasing the amount of teaching data can improve learning outcomes.

[0012] However, in the learning support device described in Patent Document 1, it is known that when the selected teaching data is incremented, even if the maximum number of epochs is reached, the learning end condition may not be met, or the learning time may become excessive.

[0013] In view of the above problems, an object of the present invention is to provide a learning device, a learning method, and a learning program that can appropriately perform model learning.

[0014] Means for solving problems

[0015] In order to achieve the above-mentioned purpose, the learning device of the present invention is characterized by a learning device that enables a model to learn, wherein the model uses teaching data for learning to calculate an evaluation value for the target data, and then compares the evaluation value with a specified threshold value and classifies it as at least one of the first label or the second label. The learning device comprises: a data acquisition unit that acquires professor candidate data that is classified as at least one of the first label or the second label; an addition processing unit that changes part of the content of the professor candidate data to generate one or more changed professor candidate data; an extraction unit that performs extraction processing, wherein the extraction processing is to combine the professor candidate data and the changed professor candidate data. The candidate data is used as a matrix, and the evaluation value is calculated for the matrix using the model, and then the data whose evaluation value falls within the specified extraction range including the threshold is extracted; a learning unit causes the model to perform learning processing at a specified learning rate; and a control unit controls the extraction unit and the learning unit; the control unit is configured to append the extracted data to the teaching data and perform the learning processing and the extraction processing when the number of extracted data extracted as a result of the extraction processing performed by the extraction unit exceeds a specified stop number, and the learning processing and the extraction processing are performed until the number of new extracted data reaches the stop number.

[0016] The learning device of the present invention includes an addition processing unit that modifies portions of candidate professor data to generate one or more modified candidate professor data. The matrix used to extract the candidate professor data includes not only the candidate professor data but also the modified candidate professor data. Because the addition of data by the addition processing unit occurs before the learning and extraction processes, learning is terminated by repeating the learning and extraction processes until the amount of extracted data reaches a specified stop count. Therefore, compared to conventional learning devices, the learning device of the present invention utilizes a larger matrix for the model, improving learning efficiency.

[0017] In the learning device of the present invention, the control unit may also be configured to, when performing extraction processing using a model that has already been learned, create a new matrix by removing the extracted data from the matrix, and then have the extraction unit perform the extraction processing on the new matrix. This configuration reduces the amount of matrix data each time the learning and extraction processes are repeated, thereby reducing the processing burden on the control unit.

[0018] The learning device of the present invention may also perform a restriction process during the extraction process, if the candidate professor data serving as the basis for change has a plurality of common candidate professor data items, to extract only a predetermined number of candidate professor data items or candidate professor data items. This configuration reduces the amount of candidate professor data items or candidate professor data items to be used during the learning process, thereby enabling rapid learning.

[0019] Furthermore, the learning device of the present invention may further include a display unit for displaying the extracted data, and a label changing unit for changing the label of the extracted data displayed on the display unit; the label can be changed by the label changing unit before the extracted data is added to the teaching data. With this configuration, the user can confirm the label determination results of the extraction unit through the label changing unit and the input unit, and can change the label through the label changing unit if the determination results are incorrect, thereby improving the accuracy of the extraction unit's determination.

[0020] In addition, in the learning device of the present invention, the control unit may also perform an expansion process to expand the amplitude of the extraction range and perform the extraction process when the number of the extracted data exceeds the stop number and does not reach the specified benchmark data number, and perform the learning process when the number of the extracted data becomes greater than or equal to the benchmark data number.

[0021] If the user changes labels using the label changing unit when the amount of extracted data is less than the reference data number, the frequency of label changes increases, placing a burden on the user. In the present invention, since the learning and extraction processes are performed only after the amount of extracted data becomes equal to or greater than the reference data number, the frequency of label selection by the user using the selection unit can be reduced.

[0022] Furthermore, in the learning device of the present invention, the control unit may also expand the width of the extraction range and reduce the learning rate as the expansion process. In this way, by expanding the extraction range, the amount of extracted data can be increased. Furthermore, the learning rate is an adjustment value for the amount of weights or biases updated within a neural network during a single learning cycle. In the expansion process of the present invention, the amount of extracted data is fine-tuned by reducing the learning rate when expanding the extraction range.

[0023] In addition, the learning device of the present invention is a learning device that enables a model to learn, wherein the model uses teaching data for learning to calculate an evaluation value for target data, and then compares the evaluation value with a specified threshold value and classifies it into at least one of the first label or the second label. The learning device comprises: a data acquisition unit that acquires teaching candidate data that is classified as at least one of the first label or the second label; an extraction unit that performs extraction processing, wherein the extraction processing is to use the teaching candidate data as a matrix and calculate an evaluation value for the matrix using the model, and then extract data whose evaluation value falls within a specified extraction range including the threshold value. data; a learning unit that causes the model to perform learning processing at a specified learning rate; and a control unit that controls the extraction unit and the learning unit; the control unit is configured to, when the number of extracted data extracted by the extraction unit exceeds a specified stop number and does not reach a specified benchmark data number, perform an expansion process to expand the amplitude of the extraction range and perform an extraction process, and when the number of the extracted data becomes greater than or equal to the benchmark data number, append the extracted data to the teaching data and perform the learning process and perform the extraction process, and perform the learning process and the extraction process until the number of new extracted data reaches the stop number.

[0024] According to the learning device constructed as described above, when the number of extractions obtained by the extraction unit is less than the specified number of reference data, the extraction range is expanded to increase the number of extractions. As the number of extractions increases, the learning process can be performed efficiently.

[0025] The learning method of the present invention is a learning method for making a model learn, wherein the model uses teaching data for learning to calculate an evaluation value for target data, and then compares the evaluation value with a specified threshold value and classifies it into at least one of the first label or the second label. The learning method includes the following steps: a data acquisition step, acquiring professor candidate data that is classified as at least one of the first label or the second label; an additional processing step, changing part of the content of the professor candidate data to generate one or more changed professor candidate data; an extraction step, using the professor candidate data and the changed professor candidate data as A matrix, and using the model to calculate an evaluation value for the matrix, and then extracting data whose evaluation value falls within a specified extraction range including the threshold; and a learning step, causing the model to learn at a specified learning rate; wherein when the number of extracted data extracted by the extraction step exceeds a specified stop number, the extracted data is appended to the teaching data to perform the learning step, and a new matrix is created by removing the extracted data from the matrix, and then the extraction step is performed on the new matrix, and the learning step and the extraction step are performed until the number of new extracted data reaches the stop number.

[0026] The learning program of the present invention is a program for causing a computer to function as each of the above-mentioned learning devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 The explanatory diagram of FIG. 1 discloses the functional structure of a learning device as an example of an embodiment of the present invention.

[0028] Figure 2 The flowchart of FIG. 1 discloses the learning method of the learning device of this embodiment.

[0029] Figure 3 The explanatory diagram of FIG. 1 shows a schematic diagram of the extraction process performed by the extraction unit of the learning device of this embodiment.

[0030] Figure 4 The explanatory figures show the display screens displayed by the display unit of the learning device of this embodiment, (A) shows the situation of good products, and (B) shows the situation of defective products.

[0031] Figure 5 The explanatory diagrams of FIG. 1 disclose the addition processing performed by the addition processing unit of the learning device of this embodiment, (A) discloses the flip processing, (B) discloses the shift processing, (C) discloses the fine rotation processing, and (D) discloses the filter processing.

[0032] Figure 6 The graph shows the relationship between the learning time and the number of data of the learning device of this embodiment. DETAILED DESCRIPTION

[0033] Next, refer to Figures 1 to 6 , describing the learning device, learning method and learning program of the embodiment of the present invention. Figure 1 The explanatory diagram of FIG. 1 discloses the functional structure of the learning device 1 of this embodiment.

[0034] The learning device 1 of this embodiment is a device that causes a model M to learn, wherein the model M calculates an evaluation value for target data and classifies it into at least a first label or a second label, so that the model M can make an appropriate judgment. Here, the term "model" refers to a computer structure that receives a certain identifiable input value, makes a certain evaluation or judgment, and then outputs an output value.

[0035] The model M has a structure including a neural network and parameters. The neural network has a structure that connects a plurality of neurons. As an example, the neural network can be a hierarchical multilayer neural network that connects a plurality of grouped neurons.

[0036] Neural networks are defined by the number of neurons and their connections. Parameters (weight coefficients, etc.) define the strength of connections between neurons or layers. In a neural network, target data is input, and the target data is evaluated and labeled based on the calculation results and parameters of multiple neurons.

[0037] Model M identifies the content of the target data and performs at least the determination of the first label or the second label. For example, if the target data is image data, the type of subject (people, vehicles, animals, etc.) or the quality of the subject (good, defective, etc.) can be used as a label. These labels are stored in association with the target data. Since the structure of model M is the same as that of Patent Document 1, a detailed description will be omitted.

[0038] Next, refer to Figure 1 The functional structure of the learning device 1 of this embodiment is described below. As a functional structure, the learning device 1 of this embodiment includes: a data acquisition unit 2, an addition processing unit 3, a control unit 4, an extraction unit 5, a learning unit 6, a display unit 7, and a label changing unit 8.

[0039] The target data used by the learning device 1 of this embodiment is professor data 11 assigned either the first label or the second label, candidate professor data 12 serving as a candidate for professor data 11, and modified candidate professor data 13 for which addition processing has been performed by the addition processing unit 3. Like the professor data 11, the candidate professor data 12 and the modified candidate professor data 13 are also assigned either the first label or the second label.

[0040] Here, the teaching data refers to data of example questions and answers presented to the model M. In addition, the teaching candidate data refers to data that can become a candidate for new teaching data given to the model M.

[0041] The data acquisition unit 2 is a functional unit that performs a data acquisition step of acquiring the candidate teaching data 12 so that the model M can be learned by the learning device 1 .

[0042] The addition processing unit 3 is a functional unit that changes part of the professor candidate data 12 to generate one or more changed professor candidate data 13. For example, when the professor candidate data 12 is image data, the following processing can be cited as the addition processing.

[0043] For example, flip processing is performed to produce data that flips the original image data left to right, flips the original image data up to down, or rotates the original image data 180 degrees; shift processing is performed to produce data that slightly moves the coordinates of the original image data in the XY direction; or slight rotation processing is performed to rotate the original image data by a slight angle to the right and left; or filter processing such as a median filter or a Gaussian filter is performed on the original image data.

[0044] The control unit 4 is a functional unit that controls the extraction unit 5 and the learning unit 6, thereby enabling the model M to effectively learn. The extraction unit 5 performs an extraction process that uses the candidate teaching data 12 and the candidate changed teaching data 13 as a matrix 14, calculates an evaluation value for this matrix 14 using the model M, and then extracts data whose evaluation value falls within a specified extraction range that includes a threshold value for the first label and the second label. The learning unit 6 performs a learning process on the model M at a specified learning rate. The learning process adjusts the parameters of the model M to approach their optimal values.

[0045] The display unit 7 is a functional unit that causes a display device such as a monitor to display data so that a user who has determined the target data can confirm the extracted data 15 extracted by the extraction unit 5. The label changing unit 8 is a functional unit that can change the label assigned to the extracted data 15 when the user confirms and decides to change the first label or the second label assigned to the extracted data 15.

[0046] The learning device 1 of this embodiment includes a computer as its main hardware (not shown). The computer includes a processor such as a CPU and GPU, RAM, ROM, a storage device such as a hard disk or SSD (solid state drive), and a communication unit connected to a network such as the Internet. The computer's storage device also stores a learning program that operates the learning device 1 of this embodiment. Furthermore, the computer includes a cloud computing system. Each of the aforementioned functional units is implemented by the computer as hardware and the learning program as software.

[0047] Next, refer to Figure 2 Taking the case where the target data is an image as an example, the learning method of the learning device 1 according to this embodiment will be described. First, the data acquisition unit 2 performs a step (STEP 1) of acquiring candidate professor data 12. This candidate professor data 12 is previously assigned either a first label or a second label. The number of candidate professor data 12 required for model M learning is prepared in advance.

[0048] Next, the addition processing unit 3 performs an addition processing step (STEP 2) to perform an addition processing on the candidate professor data 12 and generate the changed candidate professor data 13. The addition processing can be performed by any one of the aforementioned flipping, shifting, micro-rotation, and filtering processes, or a combination of these processes. The processing to be performed can be appropriately selected based on the nature of the target data, the learning objectives of the model M, or the content determined by the model M.

[0049] Next, the control unit 4 confirms whether this is the initial processing (STEP 3). If it is the initial processing (YES in STEP 3), the learning rate of the model M, which has completed learning and is in the initial state, is set to an initial value (STEP 5). The initial value of the learning rate can be set to 0.001, for example. In addition, the extraction range of the extraction unit 5 is set to an initial value (STEP 5). The initial value of the extraction range can be set to 3, for example. The content of the extraction range will be described later.

[0050] Next, the extraction unit 5 performs an extraction process (extraction step) (STEP 6). In the extraction process, the professor candidate data 12 and the changed professor candidate data 13 are used as a matrix 14. The data of the matrix 14 is input into the model M, and an evaluation value is calculated in the model M. If the evaluation value is within the extraction range, the data is extracted as extracted data 15. If the evaluation value is outside the extraction range, the data is not extracted.

[0051] Figure 3 The following schematically shows the extraction process in the learning device 1 of this embodiment. In the extraction process of the extraction unit 5, the evaluation value is calculated for each data of the matrix 14 using the model M. Figure 3 In the figure, the evaluation value is higher as you go to the right, and lower as you go to the left. In this embodiment, the one with the higher evaluation value is listed as the first label, and the one with the lower evaluation value is listed as the second label.

[0052] exist Figure 3 In , the boundary line B between the first label and the second label is the threshold that separates the two labels, and the specified range containing the threshold is used as the extraction range. Figure 3As shown in FIG, if the evaluation value is in the area close to the threshold, even if it is the first label, it is still very close to the second label, resulting in a small difference between good and bad products. Figure 3 The value of the extraction range is set to 3, but it can also be set to a value between 1 and 20. This value can be changed appropriately according to the extraction conditions or the learning object.

[0053] In this way, if model M is trained using data within the designated area of the first and second labels, it can be trained using data that is difficult to distinguish between the first and second labels. Model M trained on this data can accurately judge subtle differences. In other words, this data is excellent data that allows model M to learn efficiently.

[0054] The content of the extraction process in the extraction unit 5 is the same as the process described in Patent Document 1. Specifically, the professor candidate data 12 and the changed professor candidate data 13 are input into the model M, and based on the model M and the professor data 11, a feature quantity (vector) represented by a feature space of a predetermined dimension is calculated for each data. In addition, the distance between the professor data 11 and the professor candidate data 12 and the changed professor candidate data 13 in the feature space is calculated respectively. If the distance falls within the extraction range, the data is extracted, and if it falls outside the extraction range, the data is not extracted. In addition, since the method for calculating the distance in the feature space is the same as that in Patent Document 1, a detailed description is omitted.

[0055] In the learning device 1 of this embodiment, when the extraction processing is performed by the extraction unit 5, if the professor candidate data 12 in the mother body 14 that serves as the basis for changing the professor candidate data 13 has a plurality of common data, a limiting processing is performed to extract only a specified limit number of data in one extraction processing (STEP 6).

[0056] By performing this limiting process, when the model M is trained using the new professor data 11 in STEP 5 described later, the amount of professor data 11 does not increase rapidly, allowing for rapid learning. In this embodiment, the limited number of extractions is set to 1. Therefore, in this embodiment, even when multiple modified professor candidate data 13 are generated from the same professor candidate data 12, only one professor candidate data 12 or modified professor candidate data 13 is extracted in a single extraction process. This limited number of extractions can be appropriately changed depending on the state of the model M or the state of the hardware, etc.

[0057] Next, the control unit 4 checks the number of extracted data 15 extracted by the extraction unit 5 (STEP 7). Specifically, as (1), it is checked whether the number x of extracted data 15 exceeds the specified stop number and is greater than or equal to the specified reference data number, or whether the extraction range is an expansion range (in Figure 2 In STEP 7, "reference data number" is indicated as "reference number". Alternatively, in STEP 7, as (2), check whether the extracted data 15 exceeds the stop number and does not reach the reference data number.

[0058] Here, the stop number is a number used as a reference for whether to stop the learning performed by the learning device 1, and can be set to, for example, 0. In this case, the learning is stopped when the data extracted by the extraction unit 5 is 0.

[0059] The reference data number is used to reduce the frequency of the user's confirmation operation (STEP 8 to 9) described later. For example, the first label can be set to 20 and the second label can be set to 20. The reference data number can be appropriately changed according to factors such as the type of data and the learning environment.

[0060] In this case, if the number of extracted data 15 extracted by the extraction unit 5 is less than 20 for the first tag or less than 20 for the second tag, the user confirmation operation is not performed, and the extraction process is performed again. The user confirmation operation is not performed until the number of extracted data 15 becomes greater than or equal to 40. This process reduces the frequency of the user's confirmation operation, thereby reducing the burden on the user.

[0061] When the number of extracted data 15 exceeds the stop number and is equal to or larger than the reference data number (1 in STEP 7), the extracted data 15 is displayed on the display unit 7 (STEP 8). Figure 4 The status of the extracted data 15 is displayed on the display unit 7. On the display screen 16, a change display 18 is displayed for confirming whether the label 17 assigned to the extracted data 15 by the label changing unit 8 is appropriate and for changing the label 17.

[0062] exist Figure 4 In (A), the text "Is this a good product?" is displayed at the top of display screen 16. Extracted data 15 is displayed in the center of display screen 16. A label 17 reading "Good product" is displayed to the left of extracted data 15. Three buttons reading "Yes," "No," and "Unknown" are displayed at the bottom of display screen 16 as a change display 18. By clicking a button in change display 18, the user can confirm whether label 17 (in this case, "Good product") assigned to the extracted data is appropriate. If not, label 17 is changed.

[0063] Figure 4 (B) is a case where the label 17 is "defective product" and displays an image determined to be defective in the extracted data 15. If the user selects "unknown" in the change display 18, the user's determination is retained so that a determination can be made later.

[0064] After the user confirms the extracted data 15, the control unit 4 stores the extracted data 15 as the teaching data 11 and uses the new teaching data 11 for the next learning of the label 17. In addition, the control unit 4 removes the extracted data 15 from the matrix 14 and creates a new matrix 14 (STEP 10).

[0065] Next, it is reconfirmed whether this is the initial processing (STEP 3). Since the initial processing has already been completed and the processing is now the second or subsequent processing (NO in STEP 3), the control unit 4 performs a learning process (learning step) using the model M to which the new teaching data 11 has been added (STEP 4). Since the newly added teaching data 11 is extracted data 15 with high learning effect extracted by the extraction unit 5, the accuracy of the evaluation performed by the model M is improved. Thereafter, the learning rate is set to the initial value, and the extraction range in the extraction unit 5 is set to the initial value (STEP 5).

[0066] The control unit 4 repeats the above extraction process and learning process until the number of extracted data 15 reaches 0, which is the stop number. When the number of extracted data 15 reaches 0, which is the stop number (3 in STEP 7), the process ends.

[0067] On the other hand, if the extracted data 15 exceeds the stop count but does not reach the reference data count (2 in STEP 7), an expansion process is performed (STEP 11) to expand the extraction range (e.g., to 3.7) and reduce the learning rate (e.g., to 0.0002). If the extraction process (STEP 6) is performed in this state, the extraction range becomes wider, increasing the amount of data extracted by the extraction unit 5. Furthermore, the learning rate is fine-tuned through this extraction process.

[0068] Next, the control unit 4 checks the number of extracted data (STEP 7). In this case, since the extraction range has been expanded in STEP 11 (1 in STEP 7), the extracted data extracted by the extraction process is displayed and checked (STEPs 8 and 9). When the extraction range is expanded, the expansion process in STEP 11 has already been executed. Therefore, even if the number of extracted data 15 does not reach the reference data number, it is expected that there is still data close to the reference data number.

[0069] Next, refer to Figure 5 , a specific example of the addition processing performed by the addition processing unit 3 is described. Figure 5 (A) is an explanatory diagram illustrating the state of the flipping process. The flipping process generates vertically inverted image data, horizontally inverted image data, and 180° rotated image data from the original teaching candidate data 12. The flipping process generates three modified teaching candidate data 13 in addition to the teaching candidate data 12.

[0070] Figure 5 (B) is an explanatory diagram illustrating the shifting process. A total of four data sets—the original teaching candidate data 12 and the three modified teaching candidate data 13 generated by the flipping process—are slightly shifted in the X and Y directions by (-1, -1), (-1, 1), and so on. This shifting process yields 32 image data sets. For images, the shifting unit can be pixels, mm, or μm. The shifting range can be adjusted appropriately depending on the purpose of the learning process.

[0071] Figure 5 (C) is an explanatory diagram showing a state where a micro-rotation process has been performed, where image data rotated 1° to the left and 1° to the right are generated from the unrotated image data. The micro-rotation angle can be appropriately changed depending on the purpose of learning, etc.

[0072] Figure 5 (D) is an explanatory diagram illustrating the state of filtering. Data after applying a median filter and data after applying a Gaussian filter are generated for unfiltered image data. A median filter, also known as a median filter, is used to remove image noise. A Gaussian filter is a type of smoothing filter that blurs and smoothes an image using a Gaussian distribution. Other known filters can also be used for filtering.

[0073] If the addition process performed by the addition processing unit 3 is only Figure 5 (A) is reversed, the number of data becomes 4 times. Figure 5 The shift processing of (B) becomes 32 times the data, and if the 4 data without shift processing are added, it becomes 36 times the data. Figure 5 (C) The small rotation process, the 32-fold data that has been shifted is further multiplied by 3 times to become 96 data, and if the initial 4 data are added, it becomes 100-fold data. Figure 5 The filter processing of (D) triples the 100-fold data to 300-fold.

[0074] As described above, in the learning device 1 of this embodiment, a large amount of modified candidate teaching data 13 similar to the candidate teaching data 12 can be generated by the addition process, and therefore a large amount of data with high learning effect can be extracted by the extraction process.

[0075] Next, refer to Figure 6 The operation and effects of the learning device 1 according to this embodiment will be described. Figure 6This is a chart showing the relationship between the number of target data and the learning time. In the chart, the data shown by the solid line is the learning device 1 of this embodiment, and the data shown by the dotted line is the existing learning device (comparative example). Figure 6 As shown in the graph, the learning device 1 of this embodiment can also end learning when the target data number is 300 times. On the other hand, in the comparative example, when the increase in the target data number exceeds 8 times, learning cannot be ended.

[0076] As described above, compared to conventional learning devices, the learning device 1 of this embodiment can improve the quality of learning by increasing target data, and can still complete learning even when the target data increases.

[0077] Furthermore, in the above embodiment, when the number of extracted data 15 exceeds the stop count and is equal to or greater than the reference data count (1 in STEP 7), the extracted data 15 is displayed on the display unit 7 and the user performs a confirmation operation. However, it is also possible to add the extracted data 15 to the teaching data 11 without displaying the data on the display unit 7. For example, during the addition process, if the modified teaching candidate data 13 created based on the teaching candidate data 12 is already displayed and confirmed, the display on the display unit 7 can be omitted.

[0078] In addition, although the above embodiment describes the case where the target data is image data, the target data may also be audio data, graphic data, or video data. Furthermore, although the number of confirmed stops is set to 0 in STEP 7 in the above embodiment, this is not limited to this and may be another number (e.g., an integer such as 1 or 10).

[0079] In addition, in the above embodiment, the example of model M determining whether the subject is a good product or a defective product and being labeled as the first label or the second label is used for explanation. However, the same embodiment as above can also be applied to model M for determining the type of subject (people, vehicles, animals, etc.).

[0080] For example, it may also be: in the case where the types of the photographed objects are classified into a model of label A, label B, and label C, in STEP 7, for label A, label B, and label C, when the number of extractions of each label exceeds the benchmark data number, STEP 8 to STEP 10 are processed, and when the number of extractions of each label exceeds the stop number and does not reach the benchmark data number, STEP 11 is processed.

[0081] Furthermore, while flipping, shifting, slight rotation, and filtering are cited as examples of the added processing in the above embodiment, this is not limited to these. Various other processing methods may also be performed, such as adding impulse noise, contrast adjustment, brightness adjustment, enlargement, reduction, partial masking, cropping, deformation, or color change. Therefore, the amount of added data can be varied appropriately and is not limited to 300 times the amount in the above embodiment.

[0082] In addition, in the above embodiment, an example of using a model M that has been learned to perform learning is disclosed, but it is not limited to this, and the above process can also be performed using a model M that has not been learned. Figure 2 As the pre-processing of STEP 1, the unlearned model M is trained using the teaching data for learning, and then the processing after STEP 1 can be performed.

[0083] In the above embodiment, although in STEP 7, when the extraction range is the expanded range, the extraction data extracted by the extraction process is displayed and confirmed by moving to the side (1) (STEP 8 to 9), this is not limited to this. The extraction range may be moved to the side (2) and the extraction range may be expanded again (STEP 11) and the extraction process may be performed (STEP 6). In this case, the value of the expanded range may be changed, or the number of extraction processes (STEP 6) may be limited.

[0084] Description of Reference Numerals

[0085] M…Model

[0086] 1…Learning device

[0087] 2…Data acquisition unit

[0088] 3…Add processing unit

[0089] 4…Control Unit

[0090] 5…Extraction

[0091] 6…Study Department

[0092] 7…Display

[0093] 8…Label Change Department

[0094] 11…Teacher Data

[0095] 12…Professor candidate data

[0096] 13…Change professor candidate data

[0097] 14…Mother

[0098] 15…Extract data

[0099] 16…Display screen

[0100] 17… Tags

[0101] 18…Change display

Claims

1. A learning device, characterized in that: The learning device learns in a model, the model learning using teaching data to calculate an evaluation value for target data, then compares the evaluation value with a specified threshold and classifies the target data into at least one of a first label greater than or equal to the threshold or a second label less than the threshold, the learning device comprising: a data acquisition unit that acquires professor candidate data classified as at least one of the first label and the second label; an additional processing unit for changing part of the professor candidate data to generate one or more changed professor candidate data; an extraction unit that performs an extraction process, wherein the extraction process uses the professor candidate data and the changed professor candidate data as a matrix, calculates an evaluation value for the matrix using the model, and extracts data whose evaluation value falls within a specified extraction range including the threshold value; a learning unit that causes the model to perform a learning process at a specified learning rate, wherein the learning rate is an adjustment value for an amount by which weights or biases in a neural network are updated in one learning cycle; and a control unit, configured to control the extraction unit and the learning unit; The control unit is: When the number of extracted data extracted as a result of the extraction process performed by the extraction unit exceeds a specified stop number, the extracted data is added to the teaching data to perform the learning process and the extraction process, and the learning process and the extraction process are performed until the number of new extracted data reaches the stop number.

2. The learning device according to claim 1, wherein When performing the extraction process using the learned model, the control unit creates a new matrix by removing the extracted data from the matrix, and causes the extraction unit to perform the extraction process on the new matrix.

3. The learning device according to claim 1, wherein In the extraction process, when the professor candidate data serving as a basis for change has a plurality of the changed professor candidate data in common, a restriction process is performed in one extraction process to extract only the professor candidate data or the changed professor candidate data that is a specified limit number of extractions.

4. The learning device according to claim 1, wherein Also features: a display unit for displaying the extracted data, and a label changing unit for changing the label of the extracted data displayed on the display unit; Before the extracted data is added to the teaching data, the label may be changed by the label changing unit.

5. The learning device according to claim 4, characterized in that The control unit performs an expansion process to widen the extraction range and then performs the extraction process when the number of the extracted data exceeds the stop number and does not reach a predetermined reference number of data. When the number of the extracted data becomes equal to or larger than the number of the reference data, the learning process is performed.

6. The learning device according to claim 5, characterized in that The control unit increases the width of the extraction range and decreases the learning rate as the expansion process.

7. A learning device for enabling a model to learn, characterized in that: The model learns using teaching data to calculate an evaluation value for target data, compares the evaluation value with a specified threshold, and classifies the target data into at least one of a first label greater than or equal to the threshold or a second label less than the threshold. The learning device comprises: a data acquisition unit that acquires professor candidate data classified as at least one of the first label and the second label; an extraction unit that performs an extraction process, wherein the extraction process uses the professor candidate data as a matrix, calculates an evaluation value for the matrix using the model, and extracts data whose evaluation value falls within a specified extraction range including the threshold value; a learning unit that causes the model to perform a learning process at a specified learning rate, wherein the learning rate is an adjustment value for an amount by which weights or biases in a neural network are updated in one learning cycle; and a control unit, configured to control the extraction unit and the learning unit; The control unit is: When the number of extracted data extracted by the extraction unit exceeds the specified stop number and does not reach the specified reference data number, an expansion process is performed to expand the width of the extraction range and perform extraction processing. When the number of the extracted data becomes equal to or greater than the reference data number, the extracted data is added to the teaching data to perform the learning process and the extraction process, and the learning process and the extraction process are performed until the number of new extracted data reaches the stop number.

8. A learning method characterized by A learning method for causing a model to learn using teaching data to calculate an evaluation value for target data, then comparing the evaluation value with a specified threshold and classifying the target data into at least one of a first label greater than or equal to the threshold or a second label less than the threshold, the learning method comprising the following steps: a data acquisition step of acquiring professor candidate data classified into at least one of the first label and the second label; adding a processing step of changing part of the professor candidate data to generate one or more changed professor candidate data; an extraction step of calculating an evaluation value for the candidate professor data and the candidate changed professor data using the model as a matrix at a specified learning rate, and extracting data having the evaluation value falling within a specified extraction range including the threshold, wherein the learning rate is an adjustment value for the amount by which weights or biases in the neural network are updated during one learning cycle; and a learning step, causing the model to learn; When the number of extracted data extracted by the extraction step exceeds the specified stop number, the extracted data is added to the teaching data to perform the learning step, and a new matrix is created by removing the extracted data from the matrix, and then the extraction step is performed on the new matrix, and the learning step and the extraction step are performed until the number of new extracted data reaches the stop number.

9. A learning program product, characterized in that It is used to make a computer function as any one of claims 1 to 7 of the learning device.

Images