JP7502808B2 - Learning device, learning method, and learning program - Google Patents

Description

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

Conventionally, the results of learning through deep learning using neural networks have been used to determine the quality of products using image processing technology. In addition, in order to make learning through deep learning more efficient, Patent Document 1 proposes a learning support device and a learning device for appropriately learning a model.

The learning support device of Patent Document 1 derives features of the teacher data for each piece of teacher data based on a model trained using teacher data to classify target data into either a first label or a second label, and the teacher data having first data assigned with a first label and second data assigned with a second label.

The device also includes a derivation unit that derives features of the teacher candidate data for each teacher candidate data based on at least one of teacher candidate data to which either a first label or a second label has been assigned and a model, a calculation unit that calculates at least one of the distance between the teacher candidate data and the first data and the distance between the teacher candidate data and the second data for each teacher candidate data, and a selection unit that selects data to be added as teacher data from the teacher candidate data based on the distance.

In the learning support device of Patent Document 1, from among multiple teacher candidate data classified into a first label, data whose distance is close to the teacher data of a second label is selected, and from among multiple teacher candidate data classified into a second label, data whose distance is close to the teacher data of the first label is selected.

This process selects teacher candidate data with the first label that is close to the second label, i.e., teacher candidate data that is difficult to distinguish between the first label and the second label. The same applies to teacher candidate data with the second label. In the learning support device of Patent Document 1, by using such teacher candidate data that is difficult to distinguish as teacher data, it is possible to improve the quality of the teacher data and efficiently train the model.

JP 2021-103344 A

The learning support device described in Patent Document 1 makes it possible to carry out efficient learning. However, in general learning methods, it is generally known that the learning effect can be improved by performing an increase process in which the teacher data is increased using a specified method. For this reason, it is expected that the learning effect will also be improved by increasing the teacher data in the learning support device described in Patent Document 1.

However, when the learning support device described in Patent Document 1 performed an increase process on selected teacher data, it was found that there were cases where the learning termination condition could not be met even when the maximum number of epochs was reached, and cases where the learning time became excessive.

In view of the above problems, the present invention aims to provide a learning device, a learning method, and a learning program that can appropriately learn a model.

In order to achieve the above object, the learning device of the present invention is a learning device that performs learning on a model that uses teacher data to calculate an evaluation value for target data, compares the evaluation value with a predetermined threshold, and classifies the target data into at least one of a first label or a second label. The learning device includes a data acquisition unit that acquires teacher candidate data classified into at least one of the first label or the second label, an increase processing unit that changes part of the content of the teacher candidate data to generate one or more modified teacher candidate data, an extraction unit that performs an extraction process that uses the teacher candidate data and the modified teacher candidate data as a population, calculates an evaluation value for the population using the model, and extracts data whose evaluation value is in a predetermined extraction range including the threshold, a learning unit that causes the model to perform a learning process at a predetermined learning rate, and a control unit that controls the extraction unit and the learning unit. When the number of extracted data extracted as a result of the extraction unit performing the extraction process exceeds a predetermined stop number, the control unit adds the extracted data to the teacher data and performs the learning process, and performs the extraction process, and performs the learning process and the extraction process until the number of new extracted data reaches the stop number.

The learning device of the present invention is equipped with an augmentation processing unit that modifies part of the content of the teacher candidate data to generate one or more modified teacher candidate data, and uses not only the teacher candidate data but also the modified teacher candidate data as a population for extracting teacher candidate data. Since the data augmentation by this augmentation processing unit is performed before the learning process and the extraction process are performed, the number of extracted data eventually reaches a predetermined stopping number by repeating the learning process and the extraction process, and learning ends. Therefore, compared to conventional learning devices, the learning device of the present invention can efficiently train a model using a larger population.

In the learning device of the present invention, when performing the extraction process using a trained model, the control unit may create a new population by removing the extracted data from the population, and perform the extraction process on the new population using the extraction unit. With this configuration, the amount of data in the population decreases each time the learning process and the extraction process are repeated, thereby reducing the processing load on the control unit.

In the extraction process, when there are multiple pieces of modified teacher candidate data that share the teacher candidate data that was the basis for the change, the learning device of the present invention may perform a limiting process to extract only a predetermined limited number of teacher candidate data or modified teacher candidate data in one extraction process. With this configuration, the number of teacher candidate data or modified teacher candidate data when performing the learning process is reduced, so that the learning process can be performed quickly.

The learning device of the present invention may further include a display unit that displays the extracted data, and a label change unit that can change the label of the extracted data displayed on the display unit, and the label change unit may change the label before the extracted data is added to the teacher data. With this configuration, a user can check the label determination result in the extraction unit using the label change unit and the input unit, and if there is an error in the determination result, it can be changed by the label change unit, thereby improving the accuracy of the determination in the extraction unit.

In addition, in the learning device of the present invention, the control unit may perform the extraction process by performing an extension process to widen the width of the extraction range when the number of extracted data exceeds the stop number and is less than a predetermined reference number of data, and perform the learning process when the number of extracted data becomes equal to or greater than the reference number of data.

When the number of extracted data is less than the reference number of data, if the user changes the labels using the label change unit, the frequency of label changes increases, placing a burden on the user. In the present invention, the learning process and extraction process are performed after the number of extracted data becomes equal to or greater than the reference number of data, so that the frequency with which the user selects labels using the selection unit can be reduced.

In addition, in the learning device of the present invention, the control unit may widen the width of the extraction range and decrease the learning rate as the expansion process. In this way, by widening the extraction range, the number of extracted data can be increased. Furthermore, the learning rate is an adjustment value for the amount by which weights and biases in the neural network are updated in one learning session. In the present invention, in the expansion process, the learning rate is decreased when the extraction range is widened, thereby fine-tuning the number of extracted data.

The learning device of the present invention is a learning device that performs learning on a model that uses teacher data to calculate an evaluation value for target data, compares the evaluation value with a predetermined threshold, and classifies the target data into at least one of the first label or the second label. The learning device includes a data acquisition unit that acquires teacher candidate data classified into at least one of the first label or the second label, an extraction unit that performs an extraction process that uses the teacher candidate data as a population, calculates an evaluation value for the population using the model, and extracts data whose evaluation value is in a predetermined extraction range including the threshold, a learning unit that causes the model to perform a learning process at a predetermined learning rate, and a control unit that controls the extraction unit and the learning unit. When the number of extracted data extracted by the extraction unit exceeds a predetermined stop number and is less than a predetermined reference data number, the control unit performs an extension process that widens the width of the extraction range and performs the extraction process, and when the number of extracted data becomes equal to or greater than the reference data number, the control unit adds the extracted data to the teacher data and performs the learning process, and performs the extraction process, and performs the learning process and the extraction process until the number of new extracted data becomes the stop number.

According to the learning device configured as above, if the number of data extracted by the extraction means is less than a predetermined reference number, the width of the extraction range is widened so that the number of data extracted increases, and the number of data extracted increases, allowing the learning process to be performed efficiently.

The learning method of the present invention is a learning method for causing a model that uses teacher data to learn so as to calculate an evaluation value for target data, compare the evaluation value with a predetermined threshold value, and classify the target data into at least one of the first label or the second label, and includes a data acquisition step of acquiring teacher candidate data classified into at least one of the first label or the second label, an increase processing step of generating one or more modified teacher candidate data by modifying a part of the content of the teacher candidate data, an extraction step of calculating an evaluation value for the teacher candidate data and the modified teacher candidate data as a population using the model, and extracting data whose evaluation value is within a predetermined extraction range including the threshold value, and a learning step of causing the model to learn at a predetermined learning rate, and when the number of extracted data extracted by the extraction step exceeds a predetermined stopping number, the extracted data is added to the teacher data and the learning step is executed, and a new population is created by removing the extracted data from the population, and the extraction step is executed for the new population, and the learning step and the extraction step are executed until the number of new extracted data reaches the stopping number.

The learning program of the present invention is a program for causing a computer to function as each of the above learning devices.

FIG. 1 is an explanatory diagram showing a functional configuration of a learning device according to an embodiment of the present invention. 4 is a flowchart showing a learning method in the learning device of the present embodiment. FIG. 4 is an explanatory diagram showing an image of an extraction process performed by an extraction unit of the learning device of this embodiment. 1A and 1B are explanatory diagrams showing a display screen displayed by the display unit of the learning device of this embodiment, in which (A) shows the case of a non-defective product and (B) shows the case of a defective product. 13A to 13D are explanatory diagrams showing the increase processing by the increase processing unit of the learning device of this embodiment, where (A) shows flip processing, (B) shows shift processing, (C) shows minute rotation processing, and (D) shows filter processing. 11 is a graph showing the relationship between the learning time and the number of pieces of data in the learning device of the present embodiment.

Next, a learning device, a learning method, and a learning program according to an embodiment of the present invention will be described with reference to Figures 1 to 6. Figure 1 is an explanatory diagram showing the functional configuration of the learning device 1 according to this embodiment.

The learning device 1 of this embodiment is a device that performs learning on a model M that calculates an evaluation value for target data and classifies the data by determining at least whether it is a first label or a second label, so that the model M can make appropriate judgments. Here, a model refers to a mechanism that receives some input value that a computer can distinguish, performs some evaluation or judgment, and outputs an output value.

The model M has a structure including a neural network and parameters. The neural network has a structure in which multiple neurons are connected. As an example, the neural network can be a hierarchical multi-layer neural network in which layers are connected in which multiple neurons are grouped.

A neural network is defined by the number of neurons and their connection relationships. The strength of connections between neurons or layers is defined using parameters (such as weighting coefficients). Target data is input into a neural network, and the target data is evaluated and labeled based on the calculation results and parameters of multiple neurons.

Model M recognizes the contents of the target data and determines whether it is at least the first label or the second label. For example, if the target data is image data, the label can be the type of subject (person, vehicle, animal, etc.) or the quality of the subject (good, defective, etc.). This label is linked to the target data and stored. Note that the configuration of model M is the same as that of Patent Document 1, so a detailed description will be omitted.

Next, the functional configuration of the learning device 1 of this embodiment will be described with reference to FIG. 1. The learning device 1 of this embodiment includes, as its functional configuration, a data acquisition unit 2, an increase 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 target data handled by the learning device 1 of this embodiment are teacher data 11 to which either a first label or a second label has been assigned, teacher candidate data 12 which is a candidate for the teacher data 11, and modified teacher candidate data 13 which has been augmented by the augmentation processing unit 3. Like the teacher data 11, the teacher candidate data 12 and the modified teacher candidate data 13 are also assigned either a first label or a second label.

Here, teacher data refers to data that shows example questions and answers to be given to model M. Also, candidate teacher data refers to data that can be candidates for new teacher data to be given to model M.

The data acquisition unit 2 is a functional unit that performs a data acquisition process to acquire teacher candidate data 12 in order to learn the model M using the learning device 1.

The increase processing unit 3 is a functional unit that changes part of the content of the teacher candidate data 12 to generate one or more modified teacher candidate data 13. For example, when the teacher candidate data 12 is image data, the increase processing can be the following processing.

For example, there is flip processing, which creates data that is mirror-inverted, mirror-inverted, or rotated 180 degrees from the original image data; shift processing, which creates data in which the coordinates of the original image data are shifted slightly in the X-Y direction; micro-rotation processing, which creates data in which the original image data is rotated a small angle to the right or left; and filter processing, which applies a filter such as a median filter or Gaussian filter to the original image data.

The control unit 4 is a functional unit that controls the extraction unit 5 and the learning unit 6 to make the model M perform an effective learning process. The extraction unit 5 is a functional unit that performs an extraction process in which, using the teacher candidate data 12 and the modified teacher candidate data 13 as a population 14, an evaluation value is calculated for this population 14 using the model M, and data whose evaluation value is within a predetermined extraction range that includes the thresholds of the first label and the second label is extracted. The learning unit 6 is a functional unit that causes the model M to perform a learning process at a predetermined learning rate. The learning process is a process of adjusting the parameters of the model M so that they approach optimal values.

The display unit 7 is a functional unit that displays data on a display device such as a display so that a user who determines the target data can check the extracted data 15 extracted by the extraction unit 5. The label change unit 8 is a functional unit that can change the label attached to the extracted data 15 when the user checks and needs to change the first label or second label attached to the extracted data 15.

The learning device 1 of this embodiment 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 a RAM, a ROM, a hard disk or an SSD (solid state drive), a communication unit for connecting to a network such as the Internet, and the like. The storage device of the computer also stores a learning program for operating the computer as the learning device 1 of this embodiment. The computer includes a cloud computing system. Each of the functional units is realized by the computer as hardware and the learning program as software.

Next, referring to FIG. 2, a learning method which is the operation of the learning device 1 of this embodiment will be described using an example in which the target data is an image. First, an acquisition step is performed in which the data acquisition unit 2 acquires teacher candidate data 12 (STEP 1). This teacher candidate data 12 is previously labeled with either a first label or a second label. The number of teacher candidate data 12 required for learning the model M is prepared in advance.

Next, the augmentation processing unit 3 performs an augmentation process step (STEP 2) in which the teacher candidate data 12 is augmented to generate modified teacher candidate data 13. The augmentation process is performed by any one of the above-mentioned flip process, shift process, small rotation process, and filter process, or a combination of these processes. Which process to perform can be appropriately selected depending on the properties of the target data, the purpose of learning by the model M, or the contents of the judgment by the model M.

Next, the control unit 4 checks whether or not this is the first processing (STEP 3). If this is the first processing (YES in STEP 3), the learning rate is set to an initial value for the model M that has been trained and is in an initial state (STEP 5). The initial value of this learning rate can be set to 0.001, for example. In addition, the extraction range in the extraction unit 5 is set to an initial value (STEP 5). The initial value of this extraction range can be set to 3, for example. The contents of the extraction range will be described later.

Next, the extraction unit 5 performs an extraction process (extraction step) (STEP 6). In the extraction process, the teacher candidate data 12 and the modified teacher candidate data 13 are treated as a population 14, the data of this population 14 is input to the model M, an evaluation value is calculated in the model M, and if this evaluation value is within the extraction range, it is extracted as extracted data 15, and if the evaluation value is outside the extraction range, extraction is not performed.

Figure 3 is an image of the extraction process in the learning device 1 of this embodiment. In the extraction process in the extraction unit 5, an evaluation value is calculated for each data of the population 14 using the model M. In Figure 3, the evaluation value is higher toward the right side of the figure and lower toward the left side. In this embodiment, the higher the evaluation, the first label, and the lower the evaluation, the second label.

In FIG. 3, the boundary line B between the first label and the second label is the threshold value that separates the two labels, and a predetermined range including this threshold value is the extraction range. Here, as shown in FIG. 3, if an evaluation value is in a region close to the threshold value, it will be close to the second label even if it is the first label, and there will be little difference between good and bad products. In this embodiment, the value of the extraction range in FIG. 3 is set to 3, but it can be set to a value between 1 and 20. This value can be changed as appropriate depending on the extraction conditions and the target of learning.

In this way, by training model M using data in a specified area of the first label and the second label, it is possible to train with data that is difficult to distinguish as being the first label or the second label. Model M trained with such data will be able to accurately judge minor differences. In other words, such data is good quality data that can be used to train model M efficiently.

The content of the extraction process in the extraction unit 5 is the same as the process described in Patent Document 1. Specifically, the teacher candidate data 12 and the modified teacher candidate data 13 are input to the model M, and a feature quantity (vector) expressed in a feature space of a predetermined dimension is calculated for each data based on the model M and the teacher data 11. In addition, the distance between the teacher data 11 and the teacher candidate data 12 and the modified teacher candidate data 13 in the feature space is calculated. If this distance is within the extraction range, the data is extracted, and if it is outside the extraction range, the data is not extracted. Note that the method for calculating the distance in the feature space is the same as in Patent Document 1, so a detailed explanation will be omitted.

In the learning device 1 of this embodiment, when the extraction unit 5 performs the extraction process, if there are multiple pieces of data in the population 14 that share the same teacher candidate data 12 that was the basis for changing the changed teacher candidate data 13, a limiting process is performed to extract only a predetermined limited number of pieces of data in one extraction process (STEP 6).

By performing this limiting process, when learning the model M with new teacher data 11 in STEP 5 described below, the teacher data 11 does not increase suddenly, so that the learning process can be performed quickly. In this embodiment, the limit number of extractions is set to 1. Therefore, in this embodiment, even if multiple pieces of modified teacher candidate data 13 are generated from the same teacher candidate data 12, only one piece of teacher candidate data 12 or modified teacher candidate data 13 is extracted in one extraction process. This limit number of extractions can be changed as appropriate depending on the state of the model M, the state of the hardware, etc.

Next, the control unit 4 checks the number of extracted data 15 extracted by the extraction unit 5 (STEP 7). Specifically, (1) it is checked whether the number x of extracted data 15 exceeds a predetermined stop number and is equal to or greater than a predetermined reference data number, or whether the extraction range is an extended range (in STEP 7 of FIG. 2, the "reference data number" is written as the "reference number"). Or, in STEP 7, (2) it is checked whether the extracted data 15 exceeds the stop number but is less than the reference data number.

The stop count is a number that is a criterion for whether or not to stop learning by the learning device 1, and can be set to 0, for example. In this case, learning will be stopped when the data extracted by the extraction unit 5 becomes 0.

The standard number of pieces of data is a number for reducing the frequency of the user's confirmation work (STEPs 8-9) described below, and can be, for example, 20 for the first label and 20 for the second label. This standard number of pieces of data can be changed as appropriate depending on factors such as the type of data and the learning environment.

In this case, if the number of extracted data 15 extracted by the extraction unit 5 is less than 20 for the first label or less than 20 for the second label, the extraction process is performed again without the user's confirmation, and if the number of extracted data 15 becomes 40 or more, the user's confirmation is performed. This process reduces the frequency of the user's confirmation work, thereby reducing the burden on the user.

If the number of extracted data 15 exceeds the number of stops and is equal to or greater than the reference number of data (1 in STEP 7), the extracted data 15 is displayed on the display unit 7 (STEP 8). Figure 4 shows the state in which the extracted data 15 is displayed on the display unit 7. The display screen 16 displays a change display 18 for checking whether the label 17 attached to the extracted data 15 by the label change unit 8 is appropriate and for changing the label 17.

In FIG. 4(A), the words "Is this really a good product?" are displayed at the top of the display screen 16, the extracted data 15 is displayed in the center of the display screen 16, a label 17 is displayed as "good product" to the left of the extracted data 15, and three buttons "Yes," "No," and "Don't know" are displayed at the bottom of the display screen 16 as change indications 18. By performing an operation such as tapping the button of this change indication 18, the user can check whether the label 17 (in this case, "good product") attached to the extracted data is appropriate, and if it is not appropriate, the label 17 can be changed.

In the case of FIG. 4(B), the label 17 is "Defective" and the extracted data 15 displays an image determined to be defective. Note that when the user selects "Don't know" in the change display 18, the user's judgment is withheld and the judgment can be made later.

After the user has confirmed the extracted data 15, the control unit 4 stores the extracted data 15 as training data 11, and the new training data 11 is used for the next learning of the label 17. The control unit 4 also removes the extracted data 15 from the population 14, and creates a new population 14 (STEP 10).

Next, it is checked again whether this is the first processing or not (STEP 3), but in this case, the first processing has already ended and it is the second or subsequent processing (NO in STEP 3), so the control unit 4 performs a learning process (learning step) using the model M to which new teacher data 11 has been added (STEP 4). The newly added teacher data 11 is extracted data 15 with a high learning effect extracted by the extraction unit 5, so the accuracy of the evaluation by the model M is improved. After that, the learning rate is set to the initial value, and the extraction range in the extraction unit 5 is also set to the initial value (STEP 5).

The control unit 4 repeats the above extraction process and learning process until the number of extracted data 15 becomes 0, which is the number of stops. When the number of extracted data 15 becomes 0, which is the number of stops (STEP 7: 3), the process ends.

On the other hand, if the extracted data 15 exceeds the stop count but is less than the reference data count (2 in STEP 7), an extension process is performed to widen the extraction range to an extended range (e.g., 3.7) and lower the learning rate (e.g., 0.0002) (STEP 11). When the extraction process is performed in this state (STEP 6), the width of the extraction range is widened and the number of data extracted by the extraction unit 5 increases. This extraction process also fine-tunes the learning rate.

Next, the control unit 4 checks the number of extracted data (STEP 7), but in this case, because the extraction range has been set to an extended range in STEP 11 (1 in STEP 7), the extracted data extracted by the extraction process is displayed and confirmed (STEPs 8-9). If the extraction range is an extended range, the extension process of STEP 11 has been performed, and even if the number of extracted data 15 is less than the reference number of data, it is expected that there will be a number of data close to the reference number of data.

Next, referring to Fig. 5, a specific example of the increase processing performed by the increase processing unit 3 will be described. Fig. 5 (A) is an explanatory diagram showing the state after flip processing, in which vertically inverted image data, horizontally inverted image data, and image data rotated 180° are generated from the original image, which is the teacher candidate data 12, by the flip processing. By performing this flip processing, in addition to the teacher candidate data 12, three modified teacher candidate data 13 can be obtained.

Figure 5 (B) is an explanatory diagram showing the state after shift processing, in which the original image data, teacher candidate data 12, and the three modified teacher candidate data 13 generated by the flip processing, a total of four pieces of data, are shifted slightly in the X-Y direction in the form of (-1, -1), (-1, 1), etc. This shift processing makes it possible to obtain 32 pieces of image data. The unit of this shift may be pixels for images, or mm or μm. The range of the shift may be changed as appropriate depending on the purpose of learning, etc.

Figure 5 (C) is an explanatory diagram showing the state after micro-rotation processing, in which image data rotated 1° to the left and image data rotated 1° to the right are generated in comparison with unrotated image data. The angle of this micro-rotation can be changed as appropriate depending on the purpose of learning, etc.

Figure 5 (D) is an explanatory diagram showing the state after filter processing, in which data that has been subjected to a median filter and data that has been subjected to a Gaussian filter are generated from unfiltered image data. A median filter is also called a median filter, and is a filter used to remove noise from an image. A Gaussian filter is a type of smoothing filter that blurs and smoothes an image according to a Gaussian distribution. Note that other well-known filters may also be used for filter processing.

If the increase processing performed by the increase processing unit 3 is only the flip processing of FIG. 5(A), the amount of data will be four times as much. Furthermore, if the shift processing of FIG. 5(B) is performed, the amount of data will be 32 times as much, and if the four data pieces that are not shifted are added, the amount of data will be 36 times as much. Furthermore, if the micro-rotation processing of FIG. 5(C) is performed, the 32 times the amount of data that has been shifted is further multiplied by three to become 96 pieces of data, and if the first four pieces of data are added, the amount of data will be 100 times as much. Furthermore, if the filter processing of FIG. 5(D) is performed, the 100 times the amount of data will be tripled to become 300 times as much.

In this way, in the learning device 1 of this embodiment, a large amount of modified teacher candidate data 13 that is similar to the teacher candidate data 12 can be generated by the augmentation process, and a large amount of data with high learning effectiveness can be extracted by the extraction process.

Next, the effect of the learning device 1 of this embodiment will be described with reference to FIG. 6. FIG. 6 is a graph showing the relationship between the number of target data and learning time. In the graph, the data shown by the solid line is the learning device 1 of this embodiment, and the data shown by the dotted line is a conventional learning device (comparative example). As shown in the graph in FIG. 6, the learning device 1 of this embodiment was able to complete learning even when the number of target data was 300 times. On the other hand, in the comparative example, learning could not be completed when the increase in the number of target data exceeded 8 times.

As described above, the learning device 1 of this embodiment can improve the quality of learning by increasing the amount of target data compared to conventional learning devices, and can also terminate learning even when the amount of target data is increased.

In the above embodiment, when the number of extracted data 15 exceeds the stop number and is equal to or greater than the reference number of data (1 in STEP 7), the extracted data 15 is displayed on the display unit 7 and the user confirms it, but the extracted data 15 may be added to the teacher data 11 without displaying the data on the display unit 7. For example, in the increase process, if the modified teacher candidate data 13 created based on the teacher candidate data 12 has already been displayed and confirmed, the display on the display unit 7 can be omitted.

In the above embodiment, the target data is image data, but the target data may be audio data, graph data, video data, or other data. In the above embodiment, the number of stops to be confirmed in STEP 7 is set to 0, but this is not limited to this and may be another number (e.g., an integer such as 1 or 10).

In addition, in the above embodiment, an example is described in which model M judges the first label or the second label to determine whether the subject is good or bad, but the above embodiment can also be applied to model M that judges the type of subject (person, vehicle, animal, etc.).

For example, in the case of a model that classifies the type of subject into labels A, B, and C, in STEP 7, if the number of extractions for each of labels A, B, and C exceeds the reference number of data, the processing of STEP 8 to STEP 10 is performed, and if the number of extractions for each of the labels exceeds the stop number and is less than the reference number of data, the processing of STEP 11 is performed.

In addition, in the above embodiment, flip processing, shift processing, micro-rotation processing, and filter processing are given as examples of increasing processing, but the present invention is not limited to these, and various processing such as adding impulse noise, adjusting contrast, adjusting brightness, enlarging/reducing, partial masking, trimming, deformation, or discoloring may also be performed. Therefore, the amount of data to be increased is not limited to 300 times that of the above embodiment, and can be changed as appropriate.

In addition, in the above embodiment, an example of learning using a trained model M is shown, but this is not limiting, and the above processing may be performed using an untrained model M. In this case, as a preprocessing step of STEP 1 in FIG. 2, the untrained model M is trained using training teacher data, and then the processing from STEP 1 onward is performed.

In the above embodiment, if the extraction range is an extended range in STEP 7, the process moves to side (1) and the extracted data extracted by the extraction process is displayed and confirmed (STEPs 8-9). However, this is not limiting, and the process may move to side (2) and the extraction range may be set to an extended range again (STEP 11), and the extraction process (STEP 6) may be performed. In this case, the value of the extended range may be changed, and the number of times the extraction process (STEP 6) is performed may be limited.

M... Model 1... Learning device 2... Data acquisition unit 3... Increase processing unit 4... Control unit 5... Extraction unit 6... Learning unit 7... Display unit 8... Label change unit 11... Teacher data 12... Teacher candidate data 13... Changed teacher candidate data 14... Population 15... Extracted data 16... Display screen 17... Label 18... Changed display

Claims (9)

A learning device that performs learning on a model that uses training data to calculate an evaluation value for target data, compares the evaluation value with a predetermined threshold, and classifies the target data into at least one of a first label and a second label, the learning device comprising:
a data acquisition unit that acquires teacher candidate data classified into at least one of the first label and the second label;
an augmentation processing unit that generates one or more pieces of modified teacher candidate data by modifying a part of the content of the teacher candidate data;
an extraction unit that performs an extraction process for calculating an evaluation value for the teacher candidate data and the modified teacher candidate data as a population using the model, and extracting data in which the evaluation value is within a predetermined extraction range including the threshold value;
A learning unit that causes the model to undergo a learning process at a predetermined learning rate;
a control unit that controls 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 predetermined stop number, the extracted data is added to the teacher data and the learning process is performed while the extraction process is performed, and the learning process and the extraction process are executed until the number of new extracted data reaches the stop number. The learning device according to claim 1 ,
The control unit, when performing an extraction process using a trained model, creates a new population by removing the extracted data from the population, and performs the extraction process on the new population using the extraction unit. The learning device according to claim 1 ,
In the extraction process, when there are a plurality of changed teacher candidate data which have the same teacher candidate data that was the basis of the change, a limiting process is performed to extract only a predetermined limited number of the teacher candidate data or the changed teacher candidate data in one extraction process. The learning device according to claim 1 ,
The apparatus further includes a display unit that displays the extracted data, and a label change unit that can change the label of the extracted data displayed on the display unit,
A learning device that enables the label change unit to change the label of the extracted data before the data is added to the training data. The learning device according to claim 4,
When the number of the extracted data exceeds the stop number and is less than a predetermined reference number of data, the control unit performs an extension process to widen the width of the extraction range and performs the extraction process;
A learning device that performs the learning process when the number of the extracted data becomes equal to or greater than the reference number of data. The learning device according to claim 5,
The control unit is a learning device that performs the extension process by widening the width of the extraction range and decreasing the learning rate. A learning device that performs learning on a model that uses training data to calculate an evaluation value for target data, compares the evaluation value with a predetermined threshold, and classifies the target data into at least one of a first label and a second label, the learning device comprising:
a data acquisition unit that acquires teacher candidate data classified into at least one of the first label and the second label;
an extraction unit that performs an extraction process of calculating an evaluation value for the teacher candidate data as a population using the model and extracting data in which the evaluation value is within a predetermined extraction range including the threshold value;
A learning unit that causes the model to undergo a learning process at a predetermined learning rate;
a control unit that controls the extraction unit and the learning unit,
The control unit is
When the number of extracted data extracted by the extraction unit exceeds a predetermined number of stops and is less than a predetermined reference number of data, an extension process is performed to widen the width of the extraction range, and the extraction process is performed.
A learning device characterized in that, when the number of extracted data becomes equal to or greater than the reference number of data, the extracted data is added to the teacher data, the learning process is performed, and the extraction process is performed, and the learning process and the extraction process are performed until the number of newly extracted data becomes the stopping number. A learning method for learning a model that uses training data to calculate an evaluation value for target data, compares the evaluation value with a predetermined threshold, and classifies the target data into at least one of a first label and a second label, the learning method comprising the steps of:
a data acquisition step of acquiring teacher candidate data classified into at least either the first label or the second label;
an augmentation process for generating one or more modified teacher candidate data by modifying a part of the content of the teacher candidate data;
the teacher candidate data and the modified teacher candidate data are treated as a population, and an extraction step of calculating an evaluation value for the population at a predetermined learning rate using the model, and extracting data in which the evaluation value is within a predetermined extraction range including the threshold value; and a learning step of causing the model to learn,
A learning method characterized in that, when the number of extracted data extracted by the extraction process exceeds a predetermined stopping number, the extracted data is added to the teacher data and the learning process is executed, while a new population is created by removing the extracted data from the population, and the extraction process is executed on the new population, and the learning process and the extraction process are executed until the number of new extracted data reaches the stopping number. A learning program for causing a computer to function as a learning device according to any one of claims 1 to 7.

Images