JP6911498B2 - Learning devices, learning methods, and learning programs - Google Patents

Description

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

In recent years, machine learning such as deep learning may be used to acquire a predetermined ability. For example, Patent Document 1 discloses that a neural network device is made to acquire predetermined abilities such as identifying the type of an object appearing in an image and identifying the attribute of the person from the life log data of the person by machine learning. Has been done.

Japanese Unexamined Patent Publication No. 2016-143094

For machine learning of learners such as neural networks, learning data related to the ability to learn is used. When machine learning of the learning device is performed using a sufficient number of learning data, it is possible to construct a learned learning device that fully demonstrates the acquired ability. On the other hand, when the number of learning data used for machine learning is small, there is a possibility that a learned learning device that cannot fully demonstrate the acquired ability will be constructed. For example, when learning the ability of a learner to identify the type of an object appearing in an image, if the number of learning data used for machine learning is small, the learned learner can identify the object appearing in an image other than the learning data. The rate may be lower than the target value.

On the other hand, Patent Document 1 proposes a method of detecting whether or not learning data is insufficient in transfer learning. Specifically, it is proposed to detect under-learning depending on whether the evaluation value output from the unrelated unit deviates from the reference range when one or more unlabeled transfer destination data is input to the neural network. Has been done. However, with the method proposed in Patent Document 1, although it is possible to detect whether or not the training data is insufficient, it is not possible to specify how much the training data is insufficient.

The present invention, on the one hand, has been made in view of such circumstances, and an object of the present invention is to provide a technique capable of identifying the number of shortages of learning data used for machine learning.

The present invention employs the following configuration in order to solve the above-mentioned problems.

That is, the learning device according to one aspect of the present invention includes a learning data acquisition unit that is composed of a plurality of learning data and acquires a first learning data group for machine learning a predetermined ability by the learning device, and the learning data acquisition unit. The learning processing unit that constructs the first learning device that has learned the predetermined ability by performing machine learning of the learning device using the first learning data group, and the first learning device A deficiency number evaluation unit for evaluating the degree of deficiency of the learning data used for machine learning of the first learning device based on the result of demonstrating the predetermined ability with respect to the sample data is provided.

In this configuration, the learning data acquisition unit acquires a first learning data group used for machine learning of a predetermined ability. Then, the learning processing unit uses the acquired learning data group to perform machine learning of the learning device. As a result, it is possible to construct a first learning device that has acquired a predetermined ability. In addition, the shortage number evaluation unit evaluates the degree of shortage of the learning data used for machine learning based on the result that the learned learner exerts a predetermined ability on the sample data. That is, the degree of the number of learning data that is insufficient to construct a learned learning device that fully exerts a predetermined ability can be specified by the evaluation by the shortage number evaluation unit. Therefore, according to the configuration, it is possible to identify the number of shortages of learning data used for machine learning.

The predetermined ability may include all kinds of abilities that can be acquired by machine learning, for example, the ability to inspect an object shown in an image, the ability to control a controlled object device according to a user's preference, and the like. It may be the ability to estimate the future health state of the target person based on biological data such as blood pressure, the ability to estimate the state of the driver based on the image data of the driver, and the like. The input data to be input to the first learner when demonstrating a predetermined ability may be, for example, at least one of image data and biometric data. On the other hand, the output data output from the first learner may be data indicating an estimation result of the state of the object or the target person. The biological data is, for example, data showing at least one of heart rate, pulse rate, respiratory rate, body temperature, blood pressure, electroencephalogram, posture, and myoelectricity.

As a specific example, the predetermined capacity may be the ability to inspect an object manufactured on a production line. In this case, the input data input to the first learning device may be image data obtained by photographing the object, and the output data output from the first learning device is the inspection result of the object. It may be the data indicating. Further, the predetermined ability may be the ability to estimate the state of the driver of the vehicle. In this case, the input data to be input to the first learner may be at least one of the image data obtained by photographing the driver and the biometric data obtained from the driver, and is output from the first learner. The output data to be output may be data indicating the state of the driver. In addition, the predetermined ability may be the ability to estimate the health condition of the target person. In this case, the input data input to the first learning device may be biometric data obtained from the target person, and the output data output from the first learning device is the health state of the target person (for example, the brain). It may be data showing the possibility of developing a disease such as a vascular disease or a cardiovascular disease). Further, the sample data is, for example, data of the same type as the training data, and is data that can be used to test a predetermined ability acquired by the trained learner.

In the learning device according to the one aspect, the learning processing unit is composed of a part of a plurality of learning data constituting the first learning data group used for machine learning of the first learning device. A second learning device that has learned the predetermined ability may be further constructed by using the second learning data group. Then, the shortage number evaluation unit has a difference in the number of learning data included in each of the first learning data group and the second learning data group, and the first learning device and the second learning device, respectively. The degree of lack of learning data used for machine learning of the first learning device may be evaluated based on the difference in the results of demonstrating the predetermined ability with respect to the sample data. According to this configuration, the number of shortages of learning data used for machine learning can be identified relatively accurately.

In this configuration, a plurality of learning data groups having different numbers of learning data are used to express the influence of the difference in the number of learning data on the achievement level of machine learning. That is, by comparing the results of machine learning using each of the first learning data group and the second learning data group, the learning data that becomes the difference between the first learning data group and the second learning data group It shows the degree of influence on machine learning. Such a difference comparison can also be performed by using a third learning data group in which new learning data is added to the first learning data group. That is, in the present invention, the results of machine learning using each of the first learning data group and the third learning data group may be compared.

However, when comparing the results of machine learning using a plurality of learning data groups having different numbers of training data, it is preferable to use the second learning data group rather than using the third learning data group. This is because it is more likely that a trained learner that fully exerts a predetermined ability can be constructed more quickly by using the second training data group than by using the third learning data group. Because. That is, the second learning data group can be created by appropriately deleting the learning data from the first learning data group without performing the work of collecting new learning data. On the other hand, in order to create the third learning data, the work of collecting new learning data is performed. Therefore, when the second learning data group is used, the number of operations for collecting new learning data can be reduced as compared with the case where the third learning data group is used, and as a result, a predetermined number of times can be reduced. It is possible to quickly build a learned learner that fully demonstrates its abilities.

In the learning device according to the above aspect, the shortage number evaluation unit uses the first learning data group and the first learning data group as an index showing the relationship between the number of learning data and the degree to which the learned learning device exerts the predetermined ability. The difference in the number of learning data included in each of the second learning data groups and the difference in the result of each of the first learning device and the second learning device exerting the predetermined ability with respect to the sample data. By applying the condition, the degree of lack of learning data used for machine learning of the learned learning device may be evaluated. According to this configuration, the number of shortages of learning data used for machine learning can be identified relatively accurately.

In the learning device according to the one aspect, the shortage number evaluation unit determines the difference between the results of the first learning device and the second learning device exerting the predetermined ability with respect to the sample data. In the following cases, the learning data that is the difference between the first learning data group and the second learning data group may be evaluated as data unnecessary for machine learning of the learning device. According to this configuration, it is possible to specify whether or not unnecessary learning data is included in the learning data group used for machine learning.

In the learning device according to the one aspect, the learning data acquisition unit has a first additional learning data group composed of a plurality of learning data belonging to the first type, and a second type different from the first type. A second additional learning data group composed of a plurality of learning data belonging to the above type may be further acquired. The learning processing unit further constructs a third learning device that has learned the predetermined ability by using the learning data group and the first additional learning data group, and further constructs the learning data group and the second learning device. A fourth learning device that has learned the predetermined ability may be further constructed by using the additional learning data group of. Then, the shortage number evaluation unit is based on the difference in the results of each of the first learning device, the third learning device, and the fourth learning device exerting the predetermined ability with respect to the sample data. The type of learning data lacking in machine learning of the first learning device may be specified. According to this configuration, it is possible to identify the type of learning data that is lacking in the learning data group used for machine learning.

In the learning device according to the above aspect, the learning device may be configured by a neural network. According to this configuration, a learned learner can be easily constructed.

In the learning device according to the above aspect, when the shortage number evaluation unit inputs information indicating the number and types of learning data constituting the learning data group, the degree of the shortage of learning data used for the machine learning is determined. The degree of lack of learning data used for the machine learning may be evaluated by using a trained evaluation learner that has been trained to output the indicated information. According to this configuration, it is relatively easy to identify the number of shortages of learning data used for machine learning.

In the learning device according to the one aspect, the evaluation learning device may have been trained to further output the types of learning data lacking in the machine learning. According to this configuration, it is possible to identify the type of learning data that is lacking in the learning data group used for machine learning.

As another form of the learning device according to each of the above forms, an information processing method that realizes each of the above configurations may be used, a program may be used, or a computer or other device that records such a program. , A storage medium that can be read by a machine or the like. Here, the recording medium that can be read by a computer or the like is a medium that stores information such as a program by electrical, magnetic, optical, mechanical, or chemical action.

For example, the learning method according to one aspect of the present invention includes a step of acquiring a first learning data group in which a computer is composed of a plurality of learning data and causes a learner to perform machine learning of a predetermined ability. A step of constructing a first learning device that has learned the predetermined ability by performing machine learning of the learning device using the first learning data group, and the predetermined ability with respect to sample data. The learning data used for the machine learning based on the step of operating the first learning device so as to exhibit the above and the result of the first learning device exerting the predetermined ability with respect to the sample data. It is a step to evaluate the degree of shortage of cases and a method to execute.

Further, for example, the learning program according to one aspect of the present invention includes a step of acquiring a first learning data group for making a computer learn a predetermined ability by machine learning, which is composed of a plurality of learning data. , The step of constructing the first learning device that has learned the predetermined ability by performing machine learning of the learning device using the first learning data group, and the predetermined setting with respect to the sample data. Based on the step of operating the first learning device so as to exert the ability of the first learning device and the result of the first learning device exerting the predetermined ability with respect to the sample data, it was used for the machine learning. It is a program to evaluate the degree of lack of learning data and to execute.

According to the present invention, it is possible to provide a technique capable of identifying the number of shortages of learning data used for machine learning.

FIG. 1 schematically illustrates an example of an application scene of the learning device according to the embodiment. FIG. 2 schematically illustrates an example of the hardware configuration of the learning device according to the embodiment. FIG. 3 schematically illustrates an example of the functional configuration of the learning device according to the embodiment. FIG. 4 schematically illustrates the relationship between the learning data group and the partial learning data group according to the embodiment. FIG. 5 illustrates an example of the processing procedure of the learning device according to the embodiment. FIG. 6A shows an example of the relationship between the number of training data and the degree of ability of the trained learner. FIG. 6B shows an example of the relationship between the number of training data and the degree of ability of the trained learner. FIG. 7 schematically illustrates an example of the configuration of the learning device according to the modified example. FIG. 8 schematically illustrates an example of the configuration of the learning device according to the modified example.

Hereinafter, embodiments according to one aspect of the present invention (hereinafter, also referred to as “the present embodiment”) will be described with reference to the drawings. However, the embodiments described below are merely examples of the present invention in all respects. Needless to say, various improvements and modifications can be made without departing from the scope of the present invention. That is, in carrying out the present invention, a specific configuration according to the embodiment may be appropriately adopted. For example, in the following, as a scene in which a learner is made to acquire a predetermined ability by machine learning, a scene in which a neural network is made to acquire the ability to inspect an object flowing on a production line will be illustrated. However, the predetermined ability to be acquired by machine learning and the type of learning device are not limited to such examples, and may be appropriately selected according to the embodiment. Although the data appearing in the present embodiment is described in natural language, more specifically, it is specified in a pseudo language, a command, a parameter, a machine language, etc. that can be recognized by a computer.

§1 Application example First, an example of a situation in which the present invention is applied will be described with reference to FIG. FIG. 1 schematically illustrates an example of an application scene of the learning device 1 according to the present embodiment.

As shown in FIG. 1, the learning device 1 according to the present embodiment is an information processing device that performs machine learning of the ability to inspect an object 6 manufactured on a production line as a predetermined ability. The learning device 1 acquires, for example, a learning data group 122 composed of a plurality of learning data 123 by photographing an object 6 using a camera 5.

The learning data group 122 is a data group for causing the neural network 2, which is a learning device, to perform machine learning on the ability to inspect the object 6. The learning device 1 constructs a learned neural network 2 that has learned the ability to inspect the object 6 by performing machine learning of the neural network 2 using the acquired learning data group 122.

However, it is unclear whether the constructed trained neural network 2 can sufficiently exert the ability to inspect the object 6. In other words, it is unclear whether the constructed trained neural network 2 can correctly inspect the object 6 with a probability that it can withstand the use in the field for the data other than the training data group 122.

Therefore, the learning device 1 is made to exhibit the ability to inspect the object 6 on the sample data 129 by using the trained neural network 2. That is, the learning device 1 inputs the sample data 129 into the trained neural network 2 and operates the trained neural network 2 so as to output the inspection result of the object 6 based on the sample data 129. Then, the learning device 1 determines the degree of shortage of the learning data 123 used for machine learning based on the result of the trained neural network 2 demonstrating the ability to inspect the object 6 on the sample data 129. evaluate.

As described above, the learning device 1 acquires an evaluation for the degree of shortage of the learning data 123 used for machine learning. According to the present embodiment, based on the result of this evaluation, the degree of the number of training data 123 that is insufficient to construct a trained neural network capable of fully demonstrating the ability to inspect the object 6. Can be identified. That is, it is possible to specify the number of shortages of the learning data 123 used for machine learning.

In addition, by being able to identify the number of shortages of the training data 123, the object 6 can be inspected without adding the learning data and repeatedly performing machine learning of the neural network using the learning data group including the added learning data. You will be able to build a trained neural network that can fully demonstrate its ability to perform. Therefore, according to the present embodiment, it becomes possible to quickly create a trained neural network capable of fully exerting the ability to inspect the object 6.

§2 Configuration example [Hardware configuration]
Next, an example of the hardware configuration of the learning device 1 according to the present embodiment will be described with reference to FIG. FIG. 2 schematically illustrates an example of the hardware configuration of the learning device 1 according to the present embodiment.

As shown in FIG. 2, the learning device 1 according to the present embodiment is a computer to which the control unit 11, the storage unit 12, the external interface 13, the input device 14, the output device 15, and the drive 16 are electrically connected. .. In FIG. 2, the external interface is described as "external I / F".

The control unit 11 includes a hardware processor such as a Central Processing Unit (CPU), Random Access Memory (RAM), and Read Only Memory (ROM), and is configured to execute various types of information processing based on programs and data. NS. The storage unit 12 includes a learning program 121 executed by the control unit 11, a learning data group 122 composed of a plurality of learning data 123 used for machine learning of the neural network, and an ability acquired by the trained neural network. The sample data 129 and the like for testing the above are stored. The learning program 121 is a program for causing the learning device 1 to execute a learning process (FIG. 5) described later. Details will be described later.

The external interface 13 is an interface for connecting to an external device, and is appropriately configured according to the external device to be connected. In the present embodiment, the learning device 1 is connected to the camera 5 via the external interface 13. The camera 5 is an imaging device for monitoring an object 6 (for example, processed food, machine product, etc.) manufactured on a production line, and is appropriately arranged at a position where the object 6 can be imaged.

The input device 14 is, for example, a device for inputting a mouse, a keyboard, or the like. The output device 15 is, for example, a device for outputting a display, a speaker, or the like.

The drive 16 is, for example, a CD drive, a DVD drive, or the like, and is a drive device for reading a program stored in the storage medium 9. The type of the drive 16 may be appropriately selected according to the type of the storage medium 9. The learning program 121, the learning data group 122, and the sample data 129 may be stored in the storage medium 9.

The storage medium 9 stores the information of the program or the like by electrical, magnetic, optical, mechanical or chemical action so that the information of the program or the like recorded by the computer or other device, the machine or the like can be read. It is a medium to do. The learning device 1 may acquire at least one of the learning program 121, the learning data group 122, and the sample data 129 from the storage medium 9.

Here, in FIG. 2, as an example of the storage medium 9, a disc-type storage medium such as a CD or a DVD is illustrated. However, the type of the storage medium 9 is not limited to the disc type, and may be other than the disc type. Examples of storage media other than the disk type include semiconductor memories such as flash memories.

Regarding the specific hardware configuration of the learning device 1, the components can be omitted, replaced, or added as appropriate according to the embodiment. For example, the control unit 11 may include a plurality of hardware processors. The learning device 1 may be composed of a plurality of computers. Further, the learning device 1 may be a general-purpose server device, a personal computer (PC), or the like, in addition to an information processing device designed exclusively for the provided service.

[Functional configuration]
Next, an example of the functional configuration of the learning device 1 according to the present embodiment will be described with reference to FIG. FIG. 3 schematically illustrates an example of the functional configuration of the learning device 1 according to the present embodiment.

The control unit 11 of the learning device 1 expands the learning program 121 stored in the storage unit 12 into the RAM. Then, the control unit 11 interprets and executes the learning program 121 expanded in the RAM by the CPU to control each component. As a result, as shown in FIG. 3, the learning device 1 according to the present embodiment operates as a computer including a learning data acquisition unit 111, a learning processing unit 112, and a shortage number evaluation unit 113.

The learning data acquisition unit 111 acquires a learning data group 122 composed of a plurality of learning data 123, for causing a learner to perform machine learning of a predetermined ability. In the present embodiment, in order to make the neural network 2 acquire the ability to inspect the object 6, the learning data 123 of each case includes the image data 1231 of the object 6 flowing through the production line and the object 6. It includes inspection result data 1232 showing the inspection result. The image data 1231 is used as input data, and the inspection result data 1232 is used as teacher data.

The learning processing unit 112 constructs a learned neural network 2 that has learned the ability to inspect the object 6 by performing machine learning of the neural network 2 using the acquired learning data group 122. That is, when the learning processing unit 112 inputs the image data 1231 of each learning data 123 included in the learning data group 122, the learning processing unit 112 outputs an output value corresponding to the inspection result data 1232 paired with the input image data 1231. Train the neural network 2 to output. The trained neural network 2 constructed thereby corresponds to the "first learner" of the present invention. Further, the learning data group 122 corresponds to the "first learning data group" of the present invention.

The shortage number evaluation unit 113 lacks the learning data 123 used for machine learning of the neural network 2 based on the result that the trained neural network 2 exerts the ability to inspect the object 6 on the sample data 129. Evaluate the degree of cases. The sample data 129 is data that can be used to test the ability of the trained neural network to inspect the object 6, and may be, for example, the same type of data as the training data 123. The sample data 129 according to the present embodiment includes image data 1291 showing the object 6 flowing through the production line and inspection result data 1292 showing the result of inspecting the object 6.

The method for evaluating the degree of shortage of the learning data 123 can be appropriately set according to the embodiment. In the present embodiment, the learning processing unit 112 uses the partial learning data group 124 composed of a part of the plurality of learning data 123 constituting the learning data group 122 used for the machine learning of the neural network 2. Then, machine learning of the neural network 3 having the same configuration as the neural network 2 is performed. As a result, the learning processing unit 112 further constructs the trained neural network 3 that has learned the ability to inspect the object 6. The trained neural network 3 corresponds to the "second learner" of the present invention. Further, the partial learning data group 124 corresponds to the "second learning data group" of the present invention.

Here, the relationship between the learning data group 122 and the partial learning data group 124 will be described with reference to FIG. FIG. 4 schematically illustrates the relationship between the learning data group 122 and the partial learning data group 124. As illustrated in FIG. 4, the partial learning data group 124 is included in the learning data group 122. That is, the learning data X included in the partial learning data group 124 is also included in the learning data group 122. On the other hand, the learning data group 122 includes learning data Y that is not included in the partial learning data group 124. Therefore, the difference between the neural network 2 that uses the learning data group 122 and the neural network 3 that uses the partial learning data group 124 is learned by machine learning as to whether or not the training data Y is included. There may be a difference in the ability to inspect the object 6 to be inspected.

Therefore, the shortage number evaluation unit 113 inspects the object 6 by each of the trained neural network 2 and the neural network 3 and the difference in the number of training data 123 included in each of the training data group 122 and the partial learning data group 124. Based on the difference in the results of demonstrating the ability to perform with respect to the sample data 129, the degree of shortage of the learning data 123 used for machine learning is evaluated. That is, in the present embodiment, the difference in the number of learning data 123 (in other words, that the learning data group 122 includes the learning data Y in FIG. 4) has the ability to inspect the object 6 to be learned by machine learning. Express the impact on it. The shortage number evaluation unit 113 evaluates the degree of the shortage of the learning data 123 used for machine learning based on the degree of influence of the number of learning data 123 on the ability to inspect the object 6.

Next, each neural network (2, 3) will be described. As shown in FIG. 3, the neural network 2 is a multi-layered neural network used for so-called deep learning, and includes an input layer 21, an intermediate layer (hidden layer) 22, and an output layer 23 in order from the input. ..

In the example of FIG. 3, the neural network 2 includes the intermediate layer 22 of one layer, the output of the input layer 21 is the input of the intermediate layer 22, and the output of the intermediate layer 22 is the input of the output layer 23. ing. However, the number of intermediate layers 22 is not limited to one, and the neural network 2 may include two or more intermediate layers 22.

Each layer 21-23 comprises one or more neurons. For example, the number of neurons in the input layer 21 can be set according to the number of pixels of the input image data (for example, image data 1231). The number of neurons in the intermediate layer 22 can be appropriately set according to the embodiment. Further, the number of neurons in the output layer 23 can be set according to the number of types of inspection results of the object 6.

Neurons in adjacent layers are appropriately connected to each other, and a weight (connection load) is set for each connection based on the result of machine learning. In the example of FIG. 3, each neuron is connected to all neurons in the adjacent layer, but the connection of neurons does not have to be limited to such an example and is appropriately set according to the embodiment. You can. In addition, a threshold is set for each neuron based on the result of machine learning. Basically, the output of each neuron is determined by whether or not the sum of the products of each input and each weight exceeds the threshold value.

The neural network 3 has the same configuration as the neural network 2 (for example, the number of layers of the neural network, the number of neurons in each layer, the connection relationship between neurons, and the transfer function of each neuron are the same). That is, the neural network 3 includes an input layer 31, an intermediate layer (hidden layer) 32, and an output layer 33, and each layer 31 to 33 is configured in the same manner as each layer 21 to 23 of the neural network 2.

Each function of the learning device 1 will be described in detail in an operation example described later. In this embodiment, an example in which each function of the learning device 1 is realized by a general-purpose CPU has been described. However, some or all of the above functions may be realized by one or more dedicated hardware processors. Further, regarding the functional configuration of the learning device 1, the functions may be omitted, replaced, or added as appropriate according to the embodiment.

§3 Operation example Next, an operation example of the learning device 1 will be described with reference to FIG. FIG. 5 is a flowchart illustrating an example of the processing procedure of the learning device 1. The processing procedure described below corresponds to the "learning method" of the present invention. However, the processing procedure described below is only an example, and each processing may be changed as much as possible. Further, with respect to the processing procedure described below, steps can be omitted, replaced, and added as appropriate according to the embodiment.

(Step S101)
In step S101, the control unit 11 operates as the learning data acquisition unit 111 and acquires the learning data group 122 composed of a plurality of learning data 123.

The learning data 123 of each case can be created, for example, as follows. That is, the camera 5 captures the object 6 flowing through the production line under various shooting conditions, thereby acquiring image data of the object 6. The shooting conditions can be set by the orientation of the object 6, the posture of the object 6, the position of the object 6, the angle of illumination, the illuminance of the illumination, the position of the camera 5, the orientation of the camera 5, the magnification of the camera 5, and the like. can. Next, the object 6 reflected in the acquired image data is inspected. Then, the inspection result data showing the inspection result (for example, the result of the quality determination) of the object 6 is created, and the created inspection result data is associated with the acquired image data. As a result, the learning data 123 of each case can be created. The creation of the learning data 123 for each case may be performed manually by an operator or the like, or may be automatically performed by a robot or the like.

The sample data 129 can also be created in the same manner as the learning data 123. Therefore, a large number of sets of image data and inspection result data are created, a part of the created sets is used as learning data 123 constituting the learning data group 122, and the rest of the created sets is used as sample data 129. May be good. In addition to such a case, a part of the acquired plurality of sets may be used as the learning data 123 constituting the learning data group 122.

The control unit 11 may acquire the learning data group 122 by controlling the camera 5 and creating a plurality of learning data 123 as described above. Further, the above-mentioned learning data 123 may be created by an information processing device other than the learning device 1. In this case, the control unit 11 may acquire the learning data group 122 created by another information processing device via the network, the storage medium 9, or the like. Further, the created plurality of learning data 123 may be stored in a storage device of another information processing device such as Network Attached Storage (NAS). In this case, the control unit 11 may acquire the learning data group 122 stored in the storage device of another information processing device via the network, the storage medium 9, or the like.

(Step S102)
In the next step S102, the control unit 11 operates as the learning processing unit 112, and performs machine learning of the neural network 2 by using the learning data group 122 acquired in step S101. That is, when the image data 1231 of the training data 123 of each case is input, the neural network 2 is trained so as to output the output value corresponding to the inspection result data 1232 paired with the input image data 1231. As a result, the control unit 11 constructs a trained neural network 2 that has learned the ability to inspect the object 6.

Specifically, first, the control unit 11 prepares a neural network 2 to be subjected to learning processing. The configuration of the neural network 2 to be prepared, the initial value of the weight of the connection between each neuron, and the initial value of the threshold value of each neuron may be given by the template or by the input of the operator.

Next, the control unit 11 uses the image data 1231 of each learning data 123 included in the learning data group 122 acquired in step S101 as input data and the inspection result data 1232 as teacher data, and trains the neural network 2. I do. A gradient descent method, a stochastic gradient descent method, or the like may be used for the learning process of the neural network 2.

For example, the control unit 11 inputs the image data 1231 to the input layer 21 and determines the firing of each neuron included in each layer 21 to 23 in the forward propagation direction. As a result, the control unit 11 obtains an output value from the output layer 23 of the neural network 2. Next, the control unit 11 calculates an error between the output value obtained from the output layer 23 and the value corresponding to the inspection result data 1232 which is a pair of the input image data 1231. Subsequently, the control unit 11 calculates the error of the connection weight between each neuron and the error of each threshold value of each neuron by using the error of the output value calculated by the error back propagation method. Then, the control unit 11 updates the weight of the connection between each neuron and the value of each threshold value of each neuron based on each calculated error.

The control unit 11 repeats this series of processing for each learning data 123 included in the learning data group 122 until the output value output from the output layer 23 matches the value of the corresponding inspection result data 1232. , Perform machine learning of the neural network 2. As a result, when the image data 1231 of the training data 123 of each case is input, the trained neural network 2 (trained learning) that outputs the output value corresponding to the inspection result data 1232 paired with the input image data 1231. Can build a vessel).

(Step S103)
In the next step S103, the control unit 11 operates as the learning processing unit 112, and performs machine learning of the neural network 3 by using the partial learning data group 124. As a result, the control unit 11 constructs a trained neural network 3 (second learner) that has learned the ability to inspect the object 6.

Note that this step S103 can be performed in the same manner as in step S102, except that the partial learning data group 124 is used. The control unit 11 can create the partial learning data group 124 by appropriately thinning out the learning data 123 from the learning data group 122. The number of learning data 123 included in the partial learning data group 124 can be appropriately set according to the embodiment.

The neural network 3 before learning is initially set in the same manner as the neural network 2 described above, for example. In this step S103, when the control unit 11 inputs the image data 1231 of each learning data 123 included in the partial learning data group 124, the control unit 11 corresponds to the inspection result data 1232 paired with the input image data 1231. A trained neural network 3 that outputs an output value can be constructed.

(Step S104)
In the next step S104, the control unit 11 operates each trained neural network (2, 3) so as to exert the ability to inspect the object 6 on the sample data 129. That is, the control unit 11 inputs the sample data 129 into each trained neural network (2, 3), and outputs each trained neural network (2, 3) so as to output the inspection result of the object 6. Make it work. As a result, the control unit 11 applies the sample data 129 to each of the constructed neural networks (2, 3).

Specifically, the control unit 11 inputs the image data 1291 included in the sample data 129 into the input layer 21 of the trained neural network 2, and fires each neuron included in each layer 21 to 23 in the forward propagation direction. By making a determination, an output value is obtained from the output layer 23. When the output value obtained from the output layer 23 matches the value of the inspection result data 1292, it means that the trained neural network 2 has correctly exhibited the ability to inspect the object 6.

Therefore, the control unit 11 calculates the ratio of the output value output from the output layer 23 to the value of the inspection result data 1292 for each sample data 129, in other words, the correct answer rate of the inspection of the object 6. .. As a result, the control unit 11 obtains a result in which the trained neural network 2 exerts the ability to inspect the object 6 with respect to the sample data 129. That is, in the present embodiment, the result that the trained neural network 2 exerts the ability to inspect the object 6 with respect to the sample data 129 is the correct answer rate of the inspection of the object 6 by the trained neural network 2. Can be obtained as.

The same applies to the trained neural network 3. That is, the control unit 11 inputs the image data 1291 included in the sample data 129 into the input layer 31 of the trained neural network 3, and performs arithmetic processing in the forward propagation direction to obtain an output value obtained from the output layer 33. It is determined whether or not the value of the inspection result data 1292 matches. As a result, the control unit 11 exerts the ability of the trained neural network 3 to inspect the object 6 on the sample data 129, and as a result, the correct answer for the inspection of the object 6 by the trained neural network 3 Calculate the rate.

(Step S105)
In the next step S105, the control unit 11 functions as the shortage number evaluation unit 113, and evaluates the degree of the shortage of the learning data 123 used for machine learning of the neural network 2 based on the result obtained in step S104. ..

In the present embodiment, the control unit 11 constructs a trained neural network 2 using the training data group 122 and a trained neural network 3 using the partial learning data group 124 in the steps S102 and S103. ing. Therefore, the control unit 11 has a difference in the number of learning data 123 included in each of the learning data group 122 and the partial learning data group 124, and the correct answer rate of the inspection of the object 6 by each trained neural network (2, 3). The degree of the shortage of the learning data 123 used for machine learning may be evaluated based on the degree of difference between the two.

Whether or not the ability to inspect the object 6 can be sufficiently exerted may be determined based on the target value of the correct answer rate. That is, whether or not the control unit 11 can sufficiently exert the ability to inspect the object 6 depending on whether or not the correct answer rate of the inspection of the object 6 by the trained neural network 2 exceeds the target value. You may judge. The target value of the correct answer rate may be appropriately given by the operator's input, initial setting, or the like.

For example, the control unit 11 determines whether or not the correct answer rate of the inspection of the object 6 by the trained neural network 2 exceeds the target value. When the correct answer rate of the inspection of the object 6 by the trained neural network 2 exceeds the target value, the control unit 11 determines that the number of learning data 123 used for machine learning is sufficient. On the other hand, if the correct answer rate of the inspection of the object 6 by the trained neural network 2 does not exceed the target value, the control unit 11 determines that the number of learning data 123 used for machine learning is insufficient. , Evaluate the degree of shortage.

The method for evaluating the degree of shortage of the learning data 123 may be appropriately set according to the embodiment. For example, the control unit 11 calculates the difference between the correct answer rate of the inspection of the object 6 by the trained neural network 2 and the target value as the first difference. Further, the control unit 11 calculates the difference in the correct answer rate of the inspection of the object 6 by each learned neural network (2, 3) as the second difference. The control unit 11 applies the difference in the number of learning data 123 included in each of the learning data group 122 and the partial learning data group 124 to the ratio of the first difference and the second difference to determine the degree of the shortage. It may be specified.

Further, for example, as shown in FIGS. 6A and 6B, an index showing the relationship between the number of training data 123 and the degree of ability (correct answer rate) of the trained neural network to inspect the object 6 (hereinafter, "" (Also referred to as "evaluation index") may be given. 6A and 6B schematically illustrate an example of the evaluation index. In detail, FIG. 6A shows an example of an evaluation index that can reach the target value of the correct answer rate, in which the correct answer rate increases monotonically according to the number of learning data. On the other hand, FIG. 6B shows an example of an evaluation index in which it is difficult to reach the target value of the correct answer rate. Each evaluation index may be given as appropriate. Further, the control unit 11 may hold a plurality of evaluation indexes. Each evaluation index can be expressed by a predetermined function or the like.

The control unit 11 has a difference in the number of learning data 123 included in each of the training data group 122 and the partial learning data group 124 as the given evaluation index, and the object 6 by each trained neural network (2, 3). The difference in the correct answer rate of the inspection of the above may be applied. As a result, the control unit 11 can evaluate the degree of shortage of the learning data 123 used for machine learning.

Specifically, the control unit 11 has a difference in the number of learning data 123 included in each of the learning data group 122 and the partial learning data group 124, and inspects the object 6 by each trained neural network (2, 3). Identify the evaluation index that best matches the difference in the correct answer rate (difference in ability).

When the learning data 123 included in the learning data group 122 is data suitable for machine learning of the ability to inspect the object 6, the difference in the number of learning data 123 and the difference in the correct answer rate are shown in FIG. 6A. It is most suitable for the evaluation index as shown by. In such a case, the control unit 11 calculates the difference between the number of cases in which the target value of the correct answer rate indicated by the evaluation index can be achieved and the number of cases of learning data 123 included in the learning data group 122 as the number of insufficient cases. As a result, the control unit 11 can evaluate the degree of shortage of the learning data 123 used for machine learning.

On the other hand, when the learning data 123 included in the learning data group 122 is not appropriate data for machine learning of the ability to inspect the object 6, the difference in the number of learning data 123 and the difference in the correct answer rate are as follows. It is most suitable for the evaluation index as shown in FIG. 6B. In such a case, the control unit 11 constructs a trained neural network capable of fully demonstrating the ability to inspect the object 6 even if new training data is added to the training data group 122. Can be evaluated as difficult.

When the training data Y, which is the difference between the training data group 122 and the partial training data group 124, has little relation to the ability to inspect the object 6, each trained neural network (2, 3) has sample data 129. On the other hand, there may be a case where the degree of difference as a result of demonstrating the ability to inspect the object 6 is less than or equal to a predetermined value. For example, as shown in FIG. 6B, this corresponds to the case where the difference in the correct answer rate of the inspection of the object 6 by each trained neural network (2, 3) is not more than a predetermined value. When such a case occurs, the control unit 11 determines that the learning data Y, which is the difference between the learning data group 122 and the partial learning data group 124, is data unnecessary for machine learning of the ability to inspect the object 6. It may be evaluated as being. As a result, unnecessary learning data 123 can be reduced from the learning data group 122 used for machine learning, and the efficiency of machine learning processing can be improved.

(Step S106)
In the next step S106, the control unit 11 outputs the evaluation result performed in step S105. For example, the control unit 11 can urge the user to add the learning data 123 used for machine learning by outputting a message indicating the degree of the shortage of the learning data 123 to the output device 15 such as a display. Further, even if new learning data is added to the learning data group 122, it is difficult for the control unit 11 to construct a trained neural network capable of fully exerting the ability to inspect the object 6. By outputting the message indicating the above to the output device 15, it is possible to prompt the user to change the learning data group used for machine learning. Further, the control unit 11 outputs a message indicating that the learning data Y, which is the difference between the learning data group 122 and the partial learning data group 124, is data unnecessary for machine learning of the ability to inspect the object 6. By outputting to 15, it is possible to urge the user to delete the learning data Y. As a result, the control unit 11 ends the learning process related to this operation example.

When the trained neural network 2 capable of fully exerting the ability to inspect the object 6 can be constructed, the control unit 11 configures the trained neural network 2 and between each neuron. Information indicating the weight of the connection and the threshold value of each neuron may be stored in the storage unit 12 as learning result data. Then, the control unit 11 may appropriately transfer the created learning result data to the information processing device that inspects the object 6.

[Action / Effect]
As described above, by the processing of the learning device 1, steps S102 and S103 according to the present embodiment, the trained neural network (2, 3) is constructed by using the learning data group 122 and the partial learning data group 124, respectively. do. Then, the learning device 1 evaluates the degree of shortage of the learning data 123 used for machine learning based on the difference in the abilities of the learned neural networks (2, 3) constructed by the process of step S104. .. According to the present embodiment, based on the result of this evaluation, the degree of the number of training data 123 that is insufficient to construct a trained neural network capable of fully demonstrating the ability to inspect the object 6. Can be identified.

Further, in the present embodiment, the number of training data 123 that is insufficient to construct a trained neural network capable of fully exerting the ability to inspect the object 6 by the output processing of step S106 is shown. Can be done. As a result, the ability to inspect the object 6 can be fully exhibited without repeatedly performing the addition of the learning data 123 and the machine learning of the neural network using the learning data group including the added learning data 123. You will be able to build a trained neural network. Therefore, according to the present embodiment, it becomes possible to quickly create a trained neural network capable of fully exerting the ability to inspect the object 6.

Further, the learning device 1 according to the present embodiment evaluates the degree of the shortage of the learning data 123 by comparing the machine learning results of the learning data group 122 and the partial learning data group 124. However, the method for evaluating the degree of shortage of the learning data 123 does not have to be limited to such an example. The learning device 1 according to the present embodiment compares the results of machine learning between the additional learning data group in which the new learning data 123 is added to the learning data group 122 and the learning data group 122, and the number of shortages in the learning data 123. The degree of may be evaluated.

However, when such an evaluation method is adopted, it is preferable to use the partial learning data group 124 rather than to use the additional learning data group. This is because it is more efficient to use the partial training data group 124 than to use the additional training data group to construct a trained neural network capable of fully demonstrating the ability to inspect the object 6. This is because it can be converted. That is, the partial learning data group 124 can be appropriately created from the learning data group 122 without collecting new learning data 123. On the other hand, when the additional learning data group is created, the work process is increased by the amount of collecting new learning data 123. Therefore, in the present embodiment using the partial learning data group 124 as compared with the mode using the additional learning data group, the object 6 can suppress an increase in the number of operations for collecting new learning data 123. It will be possible to quickly create a trained neural network that can fully demonstrate the ability to perform inspections.

§4 Modifications Although the embodiments of the present invention have been described in detail above, the above description is merely an example of the present invention in all respects. Needless to say, various improvements and modifications can be made without departing from the scope of the present invention. For example, the following changes can be made. In the following, the same reference numerals will be used for the same components as those in the above embodiment, and the same points as in the above embodiment will be omitted as appropriate. The following modifications can be combined as appropriate.

<4.1>
In the above embodiment, the ability to inspect the object 6 manufactured on the production line is illustrated as a predetermined ability to be trained by the neural network. Then, the input data to be input to the trained neural network 2 when demonstrating the ability to inspect the object 6 is the image data obtained by photographing the object 6, from the trained neural network 2. The output data to be output is data indicating the inspection result of the object 6. However, the predetermined ability to train the neural network is not limited to such an example, and may be appropriately selected depending on the embodiment. By appropriately changing the training data, it is possible to appropriately change a predetermined ability to train the neural network.

For example, predetermined abilities to be trained by a neural network include the ability to control a controlled device according to the user's preference, the ability to estimate the future health state of the target person based on biological data such as blood pressure, and the image of the driver. It may be the ability to estimate the state of the driver based on the image data. In this case, the input data to be input to the neural network when demonstrating a predetermined ability may be at least one of image data and biological data, and the output data output from the neural network is an object or a target person. It may be data showing the estimation result of the state of. The biological data is, for example, data showing at least one of heart rate, pulse rate, respiratory rate, body temperature, blood pressure, electroencephalogram, posture, and myoelectricity. Such biometric data can be acquired by a biosensor such as a heart rate monitor.

As a specific example, the neural network 2 may be made to learn the ability to control the control target device according to the preference of the user as a predetermined ability. In this case, the input data included in each of the learning data and the sample data is, for example, various information used for controlling the control target device. Further, the teacher data is, for example, a control value for the controlled target device according to the user's preference in the scene indicated by various information. At this time, the input data input to the trained neural network 2 when demonstrating the ability to control the controlled target device according to the user's preference becomes various information used for controlling the controlled target device, and the trained neural network. The output data output from the network 2 becomes a control value for the controlled target device according to the user's preference.

Further, the neural network 2 may be made to learn the ability to estimate the state of the driver of the vehicle as a predetermined ability. In this case, the input data included in each of the training data and the sample data may be at least one of the image data obtained by photographing the driver and the biometric data obtained from the driver. Further, the teacher data may be data indicating the state of the driver (for example, the degree of drowsiness, the degree of concentration of driving, etc.). At this time, the input data to be input to the trained neural network 2 when demonstrating the ability to estimate the state of the driver of the vehicle is the image data obtained by photographing the driver and the biometric data obtained from the driver. At least one of them, and the output data output from the trained neural network 2 is data indicating the state of the driver.

Further, the neural network 2 may be made to learn the ability to estimate the health state of the target person as a predetermined ability. In this case, the input data included in the learning data and the sample data may be biometric data obtained from the target person. In addition, the teacher data may be data indicating the health condition of the target person (for example, the possibility of developing a disease such as cerebrovascular disease or cardiovascular disease). At this time, the input data input to the trained neural network 2 when demonstrating the ability to estimate the health state of the target person becomes the biological data obtained from the target person, and the output output from the trained neural network 2. The data is data showing the health condition of the target person.

<4.2>
Further, in the above embodiment, one trained neural network 3 is constructed by using one type of partial learning data group 124. However, the number of types of the partial learning data group 124 and the number of the trained neural networks 3 to be constructed do not have to be limited to such an example, and a plurality of cases can be obtained by using the plurality of types of the partial learning data group 124. The trained neural network 3 of the above may be constructed. That is, the control unit 11 of the learning device 1 according to the above embodiment may prepare a plurality of patterns of the partial learning data group 124. In this case, the control unit 11 constructs a plurality of learned neural networks 3 by using the prepared partial learning data group 124 of a plurality of patterns in step S103. Next, the control unit 11 applies the sample data 129 to each trained neural network 3 in step S104. Then, the control unit 11 applies the sample data 129 to each of the trained neural network 2 and the plurality of trained neural networks 3 in step S105, and based on the result, the number of shortages of the training data 123 used for machine learning. Evaluate the degree of. In this case, each partial learning data group 124 may have a different number of constituent learning data and have an inclusion relationship. That is, each partial learning data group 124 may be created by gradually reducing the learning data from the learning data group 122. As a result, the state between the learning data group 124 and the learning data group 122 having the smallest number of constituent learning data can be expressed by the other partial learning data group 124, so that a predetermined ability to be acquired by machine learning can be expressed. It is possible to appropriately express the influence of the number of training data.

<4.3>
Further, in the above embodiment, the degree of shortage of the learning data 123 used for machine learning is evaluated by using the trained neural network 3 constructed by the partial learning data group 124. However, the method for evaluating the degree of shortage of the learning data 123 does not have to be limited to such an example. For example, by applying only the correct answer rate of the inspection of the object 6 by the trained neural network 2 constructed by the learning data group 122 to the evaluation index, the control unit 11 has the number of shortages of the learning data 123 used for machine learning. The degree of may be evaluated. That is, the control unit 11 may evaluate the degree of shortage of the learning data 123 used for machine learning without using the trained neural network 3 constructed by the partial learning data group 124. In this case, step S103 may be omitted.

<4.4>
Further, as illustrated in FIG. 7, by adding learning data belonging to a predetermined type to the learning data group used for machine learning, even if the type of learning data lacking in machine learning is evaluated. good. FIG. 7 schematically illustrates the learning device 1A according to this modification. In the learning device 1A, the learning data acquisition unit 111 includes a first additional learning data group 125 composed of a plurality of learning data 126 belonging to the first type, and a second type different from the first type. A second additional learning data group 127 composed of a plurality of learning data 128 belonging to the above is further acquired. That is, in this modification, the learning data acquisition unit 111 acquires a plurality of additional learning data groups each composed of different types of learning data.

The type to which the learning data belongs can be specified by the parameters of the conditions for creating the learning data. In the case of this modification, the learning data is used for machine learning of the ability to inspect the object 6. Therefore, the type to which the learning data belongs can be specified by the change parameter of the shooting condition of the object 6. For example, in the learning data 126 belonging to the first type, the angle of the object 6 may be changed, and in the learning data 128 belonging to the second type, the brightness of the illumination may be changed. That is, the plurality of learning data 126 belonging to the first type uses the plurality of image data 1261 obtained by changing the angle of the object 6 at a predetermined pitch and keeping the other shooting conditions constant. Can be created. The plurality of learning data 128 belonging to the second type are created by using the plurality of image data 1281 obtained by changing the brightness of the illumination at a predetermined pitch and keeping the other shooting conditions constant. be able to. The inspection result data (1262, 1282) are appropriately created corresponding to the associated image data (1261, 1281) in the same manner as the learning data 123.

The learning processing unit 112 uses the learning data group 122 and the first additional learning data group 125 to perform machine learning of the neural network 40. As a result, the trained neural network 40 that has learned the ability to inspect the object 6 can be further constructed. The constructed trained neural network 40 corresponds to the "third learner" of the present invention. Further, the learning processing unit 112 uses the learning data group 122 and the second additional learning data group 127 to perform machine learning of the neural network 41. As a result, the trained neural network 41 that has learned the ability to inspect the object 6 can be further constructed. The constructed trained neural network 41 corresponds to the "fourth learner" of the present invention. Each neural network (40, 41) can be configured in the same manner as each of the above neural networks (2, 3).

In the shortage number evaluation unit 113, the difference in the results of each trained neural network (40, 41) demonstrating the ability to inspect the object 6 with respect to the sample data 129, in other words, each trained neural network. Based on the difference in the correct answer rate of the inspection of the object 6 according to (40, 41), the type of learning data lacking in machine learning is specified. For example, when the trained neural network 41 has a higher correct answer rate for the inspection of the object 6 than the trained neural network 40, the shortage number evaluation unit 113 indicates that the training data group 122 has a second It can be evaluated that the learning data belonging to the type, that is, the learning data in which the brightness of the illumination is changed is insufficient.

Except for these points, the learning device 1A is configured in the same manner as the learning device 1. In this modification, after executing the series of processes of steps S101 to S106, the control unit 11 operates as the learning data acquisition unit 111, and performs the first additional learning data group 125 and the second additional learning data group 127. You may get it. The acquisition of the first additional learning data group 125 and the second additional learning data group 127 may be performed in the above step S101. Next, the control unit 11 operates as a learning processing unit 112, and performs machine learning of the neural network 40 by using the learning data group 122 and the first additional learning data group 125. Further, the control unit 11 performs machine learning of the neural network 41 by using the learning data group 122 and the second additional learning data group 127. The machine learning process can be carried out in the same manner as in steps S102 and S103. Subsequently, the control unit 11 applies the sample data 129 to each trained neural network (40, 41) in the same manner as in step S104. Then, the control unit 11 operates as a shortage number evaluation unit 113, and based on the difference in the correct answer rate of the inspection of the object 6 by each learned neural network (40, 41), the learning lacking in machine learning Identify the type of data.

The control unit 11 may output the type of learning data lacking in the machine learning identified in this way as a message by the output device 15. As a result, it is possible to show the user the type of training data that is insufficient to construct a trained neural network capable of fully exerting the ability to inspect the object 6. Therefore, it becomes possible to efficiently create the learning data to be added, and to quickly create a trained neural network capable of fully demonstrating the ability to inspect the object 6. In particular, compared to the form in which the additional learning data group is created without limiting the type of learning data to be added, in this modified example, the learning data to be added is created as much as the type of the missing learning data can be identified. Work efficiency can be improved. The types of learning data to be added are not limited to two types. That is, the type of additional learning data may be evaluated by using three or more types of additional learning data groups.

<4.5>
Further, as illustrated in FIG. 8, an evaluation learning device such as a neural network may be used for evaluating the number of shortages of learning data in step S105. FIG. 8 schematically illustrates the learning device 1B according to this modification. The learning device 1B according to this modification is configured in the same manner as the learning device 1 except that the neural network 7 is used for evaluating the number of shortages of learning data. In this case, the control unit 11 operates as the shortage number evaluation unit 113 in step S105, and evaluates the shortage number of the learning data 123 using the neural network 7.

The neural network 7 can be configured in the same manner as each of the above neural networks (2, 3). When the neural network 7 according to this modification inputs information indicating the number and types of learning data constituting the learning data group, it outputs the insufficient number evaluation information indicating the degree of the insufficient number of learning data used for machine learning. It has been learned as. As described above, the type of training data can be specified by the shooting conditions of the image data.

The control unit 11 specifies the number and types of learning data 123 included in the learning data group 122 acquired in step S101. Then, in step S105, the control unit 11 inputs information indicating the number and types of the identified learning data 123 into the input layer of the trained neural network 7, and performs arithmetic processing in the forward propagation direction, resulting in a shortage. Obtain the output value corresponding to the number evaluation information from the output layer. With this shortage number evaluation information, it is possible to show the evaluation of the shortage number of the learning data 123. Therefore, according to this modification, it is possible to relatively easily identify the number of shortages of learning data used for machine learning.

The neural network 7 may have been trained to further output the types of training data that are lacking in machine learning. By doing so, it is possible to show the user the type of training data that is insufficient to construct a trained neural network capable of fully exerting the ability to inspect the object 6. Therefore, it becomes possible to efficiently create the learning data to be added, and to quickly create a trained neural network capable of fully demonstrating the ability to inspect the object 6.

<4.6>
Further, in the above-described embodiment and modification, a general feedforward neural network having a multi-layer structure is used as each neural network (2, 3, 40, 41, 7). However, the type of each neural network (2, 3, 40, 41, 7) does not have to be limited to such an example, and may be appropriately selected depending on the embodiment. For example, a convolutional neural network including a convolutional layer and a pooling layer may be used for each neural network (2, 3, 40, 41, 7). Further, for example, when time-series data is used as input data, each neural network (2, 3, 40, 41, 7) is connected with a recursive connection from the output side to the input side, such as from the intermediate layer to the input layer. A recursive neural network having may be used. The number of layers of each neural network (2, 3, 40, 41, 7), the number of neurons in each layer, the connection relationship between neurons, and the transfer function of each neuron are appropriately determined according to the embodiment. good.

<4.7>
Further, in the above-described embodiment and modification, each learning device and evaluation learning device are configured by a neural network. However, the type of each learner and the evaluation learner does not have to be limited to the neural network, and may be appropriately selected according to the embodiment. For each learning device and evaluation learning device, for example, a support vector machine, a self-organizing map, a learning device that learns by reinforcement learning, or the like may be used.

(Appendix 1)
With a hardware processor
A memory that holds a program to be executed by the hardware processor, and
It is a learning device equipped with
The hardware processor executes the program.
A step of acquiring a first learning data group which is composed of a plurality of learning data and causes a learner to perform machine learning of a predetermined ability, and a step of acquiring the first learning data group.
A step of constructing a first learning device that has learned the predetermined ability by performing machine learning of the learning device using the acquired first learning data group.
The step of operating the first learner so as to exert the predetermined ability with respect to the sample data, and
Based on the result of the first learning device demonstrating the predetermined ability with respect to the sample data, a step of evaluating the degree of lack of learning data used for the machine learning, and a step of evaluating the degree of shortage.
Is configured to run,
Learning device.

(Appendix 2)
The hardware processor is composed of a plurality of training data, and a step of acquiring a first training data group for machine learning a predetermined ability by a learner, and a step of acquiring the first training data group.
A step of constructing a first learning device that has learned the predetermined ability by performing machine learning of the learning device using the acquired first learning data group by a hardware processor.
A step of operating the first learner by a hardware processor so as to exert the predetermined capability with respect to the sample data.
A step of evaluating the degree of shortage of the learning data used for the machine learning based on the result that the first learning device exerts the predetermined ability with respect to the sample data by the hardware processor.
To prepare
Learning method.

1.1A / 1B ... Learning device,
11 ... Control unit, 12 ... Storage unit, 13 ... External interface,
14 ... input device, 15 ... output device, 16 ... drive,
111 ... Learning data acquisition unit, 112 ... Learning processing unit,
113 ... Insufficient number evaluation department,
121 ... Learning program,
122 ... Learning data group, 123 ... Learning data,
1231 ... Image data, 1232 ... Inspection result data,
124 ... Partial learning data group,
125 ... 1st additional training data group, 126 ... (additional) training data,
1261 ... Image data, 1262 ... Inspection result data,
127 ... 2nd additional training data group, 128 ... (additional) training data,
1281 ... Image data, 1282 ... Inspection result data,
129 ... Sample data,
1291 ... Image data, 1292 ... Inspection result data,
2 ... Neural network,
21 ... Input layer, 22 ... Intermediate layer (hidden layer), 23 ... Output layer,
3 ... Neural network,
31 ... Input layer, 32 ... Intermediate layer (hidden layer), 33 ... Output layer,
40.41 ... Neural network,
5 ... camera, 6 ... object,
7 ... Neural network for evaluation (learner for evaluation)
9 ... Storage medium

Claims (14)

A learning data acquisition unit that is composed of a plurality of learning data and acquires a first learning data group for machine learning a predetermined ability by a learning device.
A learning processing unit that constructs a first learning device that has learned the predetermined ability by performing machine learning of the learning device using the first learning data group.
A shortage number evaluation unit that evaluates the degree of shortage of learning data used for machine learning of the first learning device based on the result that the first learning device exerts the predetermined ability with respect to the sample data. When,
To prepare
Learning device. The learning processing unit uses a second learning data group composed of a part of a plurality of learning data constituting the first learning data group used for machine learning of the first learning device. Then, a second learner that has learned the predetermined ability is further constructed.
In the shortage number evaluation unit, the difference in the number of learning data included in each of the first learning data group and the second learning data group, and the first learning device and the second learning device are said to be the same. Based on the difference in the results of demonstrating the predetermined ability with respect to the sample data, the degree of lack of learning data used for machine learning of the first learning device is evaluated.
The learning device according to claim 1. The shortage number evaluation unit uses the first learning data group and the second learning data group as indexes indicating the relationship between the number of learning data and the degree to which the learned learner exerts the predetermined ability. By applying the difference in the number of learning data included and the difference in the result of each of the first learning device and the second learning device exerting the predetermined ability with respect to the sample data, the said Evaluate the degree of shortage of learning data used for machine learning of the first learning device,
The learning device according to claim 2. When the difference between the first learner and the second learner as a result of demonstrating the predetermined ability with respect to the sample data is less than or equal to the predetermined value, the shortage number evaluation unit is the first. The learning data that is the difference between the learning data group of the above and the second learning data group is evaluated as unnecessary data for machine learning of the learning device.
The learning device according to claim 2 or 3. The learning data acquisition unit includes a first additional learning data group composed of a plurality of learning data belonging to the first type, and a plurality of learning data belonging to a second type different from the first type. Further acquire the second additional learning data group composed of
The learning processing unit
Using the learning data group and the first additional learning data group, a third learning device that has learned the predetermined ability is further constructed.
Using the learning data group and the second additional learning data group, a fourth learning device that has learned the predetermined ability is further constructed.
The shortage number evaluation unit is a machine of the learned learner based on the difference in the result of each of the third learner and the fourth learner exerting the predetermined ability with respect to the sample data. Identify the types of learning data that are lacking in learning,
The learning device according to any one of claims 1 to 4. The learner is composed of a neural network.
The learning device according to any one of claims 1 to 5. When the shortage number evaluation unit inputs information indicating the number and types of learning data constituting the learning data group, the learning is performed so as to output information indicating the degree of the shortage of learning data used for the machine learning. Using the learned evaluation learner that has been performed, the degree of lack of learning data used for the machine learning is evaluated.
The learning device according to claim 1. The evaluation learner has been trained to further output the types of training data that are lacking in the machine learning.
The learning device according to claim 7. The input data to be input to the first learner when demonstrating the predetermined ability is at least one of image data and biological data, and the output data output from the first learner is a target. Data showing the estimation result of the state of an object or a target person,
The learning device according to any one of claims 1 to 8. The predetermined ability is the ability to inspect an object manufactured on a production line.
The input data to be input to the first learning device is image data obtained by photographing the object.
The output data output from the first learning device is data indicating the inspection result of the object.
The learning device according to claim 9. The predetermined ability is the ability to estimate the state of the driver of the vehicle.
The input data to be input to the first learner is at least one of the image data obtained by photographing the driver and the biometric data obtained from the driver.
The output data output from the first learner is data indicating the state of the driver.
The learning device according to claim 9. The predetermined ability is an ability to estimate the health condition of the target person, and is
The input data to be input to the first learning device is biometric data obtained from the target person, and is
The output data output from the first learning device is data indicating the health state of the target person.
The learning device according to claim 9. The computer
A step of acquiring a first learning data group which is composed of a plurality of learning data and causes a learner to perform machine learning of a predetermined ability, and a step of acquiring the first learning data group.
A step of constructing a first learning device that has learned the predetermined ability by performing machine learning of the learning device using the first learning data group.
The step of operating the first learner so as to exert the predetermined ability with respect to the sample data, and
Based on the result of the first learning device demonstrating the predetermined ability with respect to the sample data, a step of evaluating the degree of lack of learning data used for the machine learning, and a step of evaluating the degree of shortage.
To execute,
Learning method. On the computer
A step of acquiring a first learning data group which is composed of a plurality of learning data and causes a learner to perform machine learning of a predetermined ability, and a step of acquiring the first learning data group.
A step of constructing a first learning device that has learned the predetermined ability by performing machine learning of the learning device using the first learning data group.
The step of operating the first learner so as to exert the predetermined ability with respect to the sample data, and
Based on the result of the first learning device demonstrating the predetermined ability with respect to the sample data, a step of evaluating the degree of lack of learning data used for the machine learning, and a step of evaluating the degree of shortage.
A learning program to execute.

Images