JP7043742B2 - Learning model creation device - Google Patents

Description

The present invention relates to a learning model creation device that creates a learning model for machine learning.

Machine learning automatically generates a learning model by iteratively learning the learning data, and the learning data can be classified using this generated learning model, or the classification of unknown data can be performed. Since it can be predicted, it has been applied to various fields and is being researched and developed. In particular, in recent years, deep learning (deep learning) represented by a convolutional neural network (hereinafter, abbreviated as "CNN") has attracted attention because of its high recognition accuracy of image recognition. Has been researched and developed.

In such machine learning, the generation of a learning model requires a huge calculation cost (calculation cost) because the training data is repeatedly learned. Further, even when analyzing unknown data using this generated learning model, if the learning model is large, the calculation cost will increase accordingly. In particular, in recent years, the number of layers of learning models in deep learning has increased, and learning models with 1000 layers have appeared. Therefore, for example, Patent Document 1 discloses a method of optimizing a learning model by removing redundant information of the learning model.

The neural network optimization method disclosed in Patent Document 1 is a method of optimizing the structure of a neural network, and includes (1) a step of inputting an initial structure of the neural network as a first neural network, and (2). ) A step of learning about a given first neural network using training data, until the cost of the first neural network calculated using the evaluation data becomes the minimum first cost. Steps to generate a second neural network by randomly deleting units from the first neural network, and (4) learning about the second neural network using training data. A step of learning until the cost of the second neural network calculated using the evaluation data becomes the minimum second cost, and (5) the first cost and the second cost. And (6) when the second cost is smaller than the first cost, the second neural network is referred to as the first neural network, and the second cost is referred to as the first cost. Steps (3) to (5) are performed as the cost of, and when the first cost is smaller than the second cost, a different second neural network is generated in step (3) to generate steps (4) and (5). ) And the step (7) (6), when it is determined that the first cost is smaller than the second cost a predetermined number of times, the first neural network is used as a neural network. It is provided with a step of determining the optimum structure of the above and (8) a step of outputting the optimum structure of the neural network. According to Patent Document 1, this neural network optimization method can automatically determine a network structure with improved generalization ability and reduced computational complexity.

Japanese Unexamined Patent Publication No. 2015-11510 (Patent No. 6042274)

By the way, in the neural network optimization method disclosed in Patent Document 1, since a neural network (an example of a learning model) is optimized, the calculation cost is reduced when this neural network is used. However, the neural network optimization method disclosed in Patent Document 1 has learning steps (2) and (4) of generating first and second neural networks in order to generate an optimized neural network. It is necessary to carry out a plurality of times from step (7), and the calculation cost cannot be reduced in the generation of the optimized neural network. Further, in the neural network optimization method disclosed in Patent Document 1, the deletion target is randomly determined, and the evaluation of the second neural network is performed by comparing the first cost and the second cost. There is a possibility that the deletion target contributes to the performance of the neural network, and the performance of the optimized neural network may be deteriorated as compared with that before the optimization.

The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a learning model creating device capable of reducing a calculation cost while suppressing performance deterioration of a learning model.

As a result of various studies, the present inventor has found that the above object can be achieved by the following invention. That is, the learning model creating device according to one aspect of the present invention is a learning model creating device that generates a learning model for machine learning, and the learning model acquires a function of extracting a predetermined feature amount by learning. Equipped with a feature quantity extraction unit of A model improvement unit that creates a new learning model by deleting the feature quantity extraction unit specified by the specific unit from the learning model, and a new learning model created by the model improvement unit. The learning application unit for continuing the learning or analyzing the input sample is provided, and the specific unit identifies the redundant feature quantity extraction unit with priority on maintaining the performance of the learning model, and the specific unit. Specifies the redundant feature amount extraction unit in order from the feature amount extraction unit having a low degree of independence indicating the degree of independence between the feature amount extraction units. Preferably, the above-mentioned learning model creating device further includes an input unit that receives an input of a specific condition used when specifying the redundant feature amount extraction unit, and the specific unit extracts the feature amount with a low degree of independence. In order from the unit, the redundant feature amount extraction unit is specified while the specific conditions accepted by the input unit are satisfied. Preferably, in the above-mentioned learning model creating device, the specific unit is the remainder of the feature quantity extraction unit and the plurality of feature quantity extraction units excluding the feature quantity extraction unit for each of the plurality of feature quantity extraction units. The difference between the components with each feature extraction unit is obtained, the sum of the squares of the obtained differences is obtained as the degree of independence, and based on the obtained degree of independence, the feature amount extraction unit having a low degree of independence is used. In order, the redundant feature amount extraction unit is specified. Preferably, the above-mentioned learning model creating device further includes an input unit that receives an input of a specific condition used when specifying the redundant feature amount extraction unit, and the specific unit is each of the plurality of feature amount extraction units. The difference between the components is obtained between the feature amount extraction unit and each of the remaining feature amount extraction units excluding the feature amount extraction unit from the plurality of feature amount extraction units, and the sum of the squares of the obtained differences is calculated. Is obtained as the degree of independence, and based on the obtained degree of independence, the redundant feature amount extraction unit is used in order from the feature amount extraction unit having the lowest degree of independence while the specific conditions accepted by the input unit are satisfied. Identify. Preferably, the specific condition is the number of deletion settings, which is the number of feature amount extraction units specified as the redundant feature amount extraction unit. Here, the term "redundancy" means that the feature amount extraction unit extracts substantially the same feature amount .

Such a learning model creating device can appropriately identify the redundant feature amount extraction unit from the viewpoint of the degree of independence. In the learning model creating device, since it is specified from the viewpoint of the degree of independence, the redundant feature amount extraction unit is specified with priority given to reduction of calculation cost.

The learning model creating device according to another aspect of the present invention is a learning model creating device that generates a learning model for machine learning, and the learning model acquires a function of extracting a predetermined feature amount by learning. The feature amount extraction unit is provided, and a redundant feature amount extraction unit is specified from the plurality of feature amount extraction units during training or when an input sample is analyzed using a learning model after learning. A model improvement unit that creates a new learning model by deleting the feature amount extraction unit specified by the specific unit from the learning model, and a new learning model created by the model improvement unit. It is provided with a learning application unit for continuing the learning or analyzing the input sample, and the specific unit identifies the redundant feature amount extraction unit and outputs color information in order to prioritize the maintenance of the performance of the learning model. From the first feature amount extraction unit that extracts the feature amount, the second feature amount extraction unit that extracts the edge information as the feature amount, and the third feature amount extraction unit that extracts the blur information as the feature amount. Further, a first input unit for receiving the input of the feature amount extraction unit to be excluded from the redundant feature amount extraction unit to be specified is further provided, and the specific unit corresponds to the feature amount extraction unit received by the first input unit. The redundant feature amount extraction unit is specified by excluding the quantity extraction unit from the redundant feature amount extraction unit to be specified.

When such a learning model creating device accepts the first input unit to exclude the first feature amount extraction unit, the first feature amount extraction unit is excluded and the redundant feature amount extraction unit is specified. Therefore, the first feature amount extraction unit can be left in the learning model after the redundant feature amount extraction unit is deleted. In this case, since the first feature amount extraction unit is left in the learning model creation device, the redundant feature amount extraction unit is specified with priority given to maintaining the performance of the learning model. When the first input unit accepts to exclude the second feature amount extraction unit, the learning model creation device excludes the second feature amount extraction unit and specifies the redundant feature amount extraction unit. The second feature amount extraction unit can be left in the learning model after the redundant feature amount extraction unit is deleted. In this case, since the second feature amount extraction unit is left in the learning model creation device, the redundant feature amount extraction unit is specified with priority given to maintaining the performance of the learning model. When the first input unit accepts to exclude the third feature amount extraction unit, the learning model creation device excludes the third feature amount extraction unit and specifies the redundant feature amount extraction unit. , The third feature amount extraction unit can be left in the learning model after the redundant feature amount extraction unit is deleted. In this case, since the learning model creating device leaves the third feature amount extraction unit, the redundant feature amount extraction unit is specified with priority given to maintaining the performance of the learning model.

The learning model creating device according to another aspect of the present invention is a learning model creating device that generates a learning model for machine learning, and the learning model acquires a function of extracting a predetermined feature amount by learning. The feature amount extraction unit is provided, and a redundant feature amount extraction unit is specified from the plurality of feature amount extraction units during training or when an input sample is analyzed using a learning model after learning. A model improvement unit that creates a new learning model by deleting the feature amount extraction unit specified by the specific unit from the learning model, and a new learning model created by the model improvement unit. It is provided with a learning application unit for continuing the learning or analyzing the input sample, and the specific unit identifies the redundant feature amount extraction unit and outputs color information in order to prioritize the maintenance of the performance of the learning model. From the first feature amount extraction unit that extracts the feature amount, the second feature amount extraction unit that extracts the edge information as the feature amount, and the third feature amount extraction unit that extracts the blur information as the feature amount. Further, a second input unit that receives the input of the feature amount extraction unit that is preferentially specified as the redundant feature amount extraction unit is further provided, and the specific unit corresponds to the feature amount extraction unit that is received by the second input unit. The redundant feature amount extraction unit is specified by preferentially specifying the quantity extraction unit as the redundant feature amount extraction unit.

When such a learning model creating device receives the second input unit to preferentially specify the first feature amount extraction unit, the first feature amount extraction unit is prioritized and the redundant feature amount extraction unit is extracted. Since the unit is specified, the first feature quantity extraction unit can be excluded from the learning model after the redundant feature quantity extraction unit is deleted. In particular, when the first feature quantity extraction unit becomes noise from the viewpoint of the usage of the learning model, the learning model creation device becomes noise from the learning model after the redundant feature quantity extraction unit is deleted. The feature amount extraction unit can be deleted (excluded). In this case, in the learning model creating device, the first feature amount extraction unit is preferentially deleted, so that the redundant feature amount extraction unit is specified with priority given to maintaining the performance of the learning model. When the second input unit accepts to preferentially specify the second feature amount extraction unit, the learning model creation device gives priority to the second feature amount extraction unit and selects the redundant feature amount extraction unit. Since it is specified, the second feature amount extraction unit can be excluded from the learning model after the redundant feature amount extraction unit is deleted. In particular, when the second feature amount extraction unit becomes noise from the viewpoint of the usage of the learning model, the learning model creating device becomes noise from the learning model after the redundant feature amount extraction unit is deleted. The feature amount extraction unit can be deleted. In this case, in the learning model creating device, the second feature amount extraction unit is preferentially deleted, so that the redundant feature amount extraction unit is specified with priority given to maintaining the performance of the learning model. When the second input unit accepts to preferentially specify the third feature amount extraction unit, the learning model creation device gives priority to the third feature amount extraction unit and selects the redundant feature amount extraction unit. Since it is specified, the learning model after deleting the redundant feature amount extraction unit can be made to not include the third feature amount extraction unit. In particular, when the third feature amount extraction unit becomes noise from the viewpoint of the usage of the learning model, the learning model creating device becomes noise from the learning model after the redundant feature amount extraction unit is deleted. The feature amount extraction unit can be deleted. In this case, in the learning model creating device, the third feature amount extraction unit is preferentially deleted, so that the redundant feature amount extraction unit is specified with priority given to maintaining the performance of the learning model.

The learning model creating device according to the present invention can reduce the calculation cost while suppressing the deterioration of the performance of the learning model.

It is a block diagram which shows the structure of the learning model making apparatus in 1st to 5th Embodiment. It is a block diagram which shows one structural example of the learning application part in the said learning model creation apparatus. It is a block diagram which shows the structure of the computer which implemented the said learning model creation apparatus. It is a flowchart which shows the operation of the learning model creation apparatus in learning. It is a flowchart which shows the operation of the learning model creation apparatus in analysis. It is a flowchart which shows the redundant deletion process in each flowchart shown in FIG. 4 and FIG. 5, respectively. It is a figure which shows the relationship between the number of repetitions of machine learning, and the total amount of information of a feature amount extraction part. This is for explaining the redundant deletion process based on the contribution in the first embodiment. It is a figure for demonstrating the effect of the learning model creation apparatus in 1st Embodiment. This is for explaining the redundant deletion process by the sum of squared differences in the second embodiment. This is for explaining the redundant deletion process by dimensional compression in the third embodiment. It is a figure which shows an example of the input / output screen displayed on the learning model creation apparatus in 4th Embodiment. It is a figure for demonstrating the effect of the learning model making apparatus in 4th Embodiment.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings. It should be noted that the configurations with the same reference numerals in the respective drawings indicate the same configurations, and the description thereof will be omitted as appropriate. In the present specification, when generically referred to, they are indicated by reference numerals without subscripts, and when referring to individual configurations, they are indicated by reference numerals with subscripts.

(First Embodiment)
FIG. 1 is a block diagram showing a configuration of a learning model creating device according to the first embodiment. Note that FIG. 1 also shows the configuration of the learning model creating device according to the second to fifth embodiments. FIG. 2 is a block diagram showing a configuration example of a learning application unit in the learning model creation device. FIG. 3 is a block diagram showing a configuration of a computer on which the learning model creation device is mounted.

The learning model creation device Da in the first embodiment is a device that generates a learning model for machine learning, and the learning model includes a plurality of feature quantity extraction units that acquire a function of extracting a predetermined feature quantity by learning. Be prepared. As shown in FIG. 1, for example, the learning model creating device Da in the first embodiment has an input unit 1, an output unit 2, an interface unit (IF unit) 3, a control processing unit 4a, and a storage unit. A unit 5 is provided.

The input unit 1 is connected to the control processing unit 4a, and for example, various commands such as a command for instructing the start of learning and a command for instructing the analysis of the input sample (target sample) using the learning model generated by the learning. And, for example, it is a device that inputs various data necessary for generating a learning model, such as input of a specific condition used when specifying the redundant feature amount extraction unit, to the learning model creation device Da, for example. A plurality of input switches, keyboards, mice, etc. to which predetermined functions are assigned.

The output unit 2 is connected to the control processing unit 4a, and is analyzed using commands and data input from the input unit 1 and a learning model generated by the learning model creating device Da according to the control of the control processing unit 4a. It is a device that outputs results and the like, and is, for example, a display device such as a CRT display, an LCD (liquid crystal display) and an organic EL display, and a printing device such as a printer.

A touch panel may be configured from the input unit 1 and the output unit 2. In the case of configuring this touch panel, the input unit 1 is a position input device that detects and inputs an operation position such as a resistance film method or a capacitance method, and the output unit 2 is a display device. In this touch panel, a position input device is provided on the display surface of the display device, one or more input content candidates that can be input to the display device are displayed, and the user touches the display position displaying the input content to be input. Then, the position is detected by the position input device, and the display content displayed at the detected position is input to the learning model creation device Da as the operation input content of the user. With such a touch panel, since the user can intuitively understand the input operation, the learning model creation device Da that is easy for the user to handle is provided.

The IF unit 3 is, for example, a circuit for inputting / outputting data to / from an external device, for example, an interface circuit of RS-232C which is a serial communication method, an interface circuit using the Bluetooth (registered trademark) standard, and IrDA. An interface circuit for performing infrared communication such as the (Infrared Data Association) standard, and an interface circuit using the USB (Universal Serial Bus) standard. The IF unit 3 may be a circuit that communicates with an external device via a communication network (network), and may be, for example, a data communication card, a communication interface circuit according to the IEEE802.11 standard, or the like. May be.

The storage unit 5 is a circuit connected to the control processing unit 4a and stores various predetermined programs and various predetermined data under the control of the control processing unit 4a. The various predetermined programs include, for example, a control program that controls each part 1 to 3 and 5 of the learning model creation device Da according to the function of each part, and learning during or after learning. When analyzing an input sample using a model, a specific program that specifies a redundant feature amount extraction unit from the plurality of feature amount extraction units, or a feature amount extraction unit specified by the specific program from the learning model. Control of a learning application program that continues the learning or analyzes the input sample with a model improvement program that creates a new learning model by deleting the above, or a new learning model created by the model improvement program. A processing program is included. The various predetermined data include data necessary for executing each program, such as a teacher set used for machine learning and an input sample (target sample) to be analyzed. The storage unit 5 includes, for example, a ROM (Read Only Memory) which is a non-volatile storage element, an EEPROM (Electrically Erasable Programmable Read Only Memory) which is a rewritable non-volatile storage element, and the like. The storage unit 5 includes a RAM (Random Access Memory) or the like that serves as a working memory of the so-called control processing unit 4a that stores data or the like generated during the execution of the predetermined program. Then, the storage unit 5 functionally includes a teacher set storage unit 51 and a target sample storage unit 52 in order to store the teacher set and the target sample.

The teacher set storage unit 51 stores the teacher set. The teacher set is learning data whose correct answer is known in advance for creating a learning model by machine learning, and is associated with teacher data composed of a plurality of samples (data) and each of the plurality of samples in the teacher data. , A correct answer label composed of a plurality of correct answer information representing the correct answer of the sample. Correspondingly, the teacher set storage unit 51 functionally includes a teacher data storage unit 511 for storing the teacher data and a correct answer label storage unit 512 for storing the correct answer label. The target sample storage unit 52 stores the target sample. The teacher set is input from the input unit 1, for example, and stored in the teacher set storage unit 51. Further, for example, it is input via the IF unit 3 from a server device that manages the teacher set, or from a recording medium such as a USB memory that stores (records) the teacher set, and is stored in the teacher set storage unit 51. To. Similarly, the target sample is input from, for example, the input unit 1 and stored in the target sample storage unit 52. Further, for example, it is input via the IF unit 3 from a server device that manages the target sample, or from a recording medium such as a USB memory that stores (records) the target sample, and is stored in the target sample storage unit 52. To.

The control processing unit 4a is a circuit for creating a learning model for machine learning by controlling each unit 1 to 3 and 5 of the learning model creating device Da according to the function of each unit. The control processing unit 4a is configured to include, for example, a CPU (Central Processing Unit) and peripheral circuits thereof. The control processing unit 4a functionally includes a control unit 41, a learning application unit 42, a specific unit 43a, and a model improvement unit 44 by executing the control processing program.

The control unit 41 controls each unit 1 to 3 and 5 of the learning model creation device Da according to the function of each unit, and controls the entire learning model creation device Da.

The specific unit 43a is a redundant feature amount from a plurality of feature amount extraction units in the learning model when the input sample (target sample) is analyzed during learning or by using the learning model after learning. It specifies the extraction unit. In the present embodiment, the specific unit 43a specifies the redundant feature amount extraction unit with priority given to reduction of calculation cost. More specifically, the specific unit 43a identifies the redundant feature amount extraction unit in order from the feature amount extraction unit having a low contribution degree indicating the degree of contribution to the performance of the learning model. More specifically, the specific unit 43a first creates a learning model for trial by deleting one feature amount extraction unit from the plurality of feature amount extraction units in the learning model. Next, the specific unit 43a analyzes (classifies) the teacher data using this learning model for trial. Next, the specific unit 43a deletes the evaluation value for evaluating the performance of the learning model for trial by collating the analysis result (classification result) with the correct answer label corresponding to the teacher data. It is calculated as the degree of contribution to the feature amount extraction unit. The specific unit 43a performs each process of creating a learning model for such a trial, analysis using the learning model, and calculation of the contribution degree for each of the plurality of feature quantity extraction units in the learning model, and the plurality of feature quantities. Obtain each contribution of each extraction unit. Then, the specific unit 43a arranges the plurality of feature quantity extraction units in order from the one with the lowest contribution, and the feature quantity extraction units are arranged in order from the lowest to the preset number (the number to be deleted to be described later). It is selected, and the selected feature amount extraction unit is specified as the redundant feature amount extraction unit.

The model improvement unit 44 creates a new learning model by deleting the feature amount extraction unit specified by the specific unit 43a from the learning model.

The learning application unit 42 is a new learning model created by the model improvement unit 44, and continues learning or analyzes an input sample. The learning application unit 42 may be any system as long as it is a machine learning system using a learning model including a plurality of feature amount extraction units that acquire a function of extracting a predetermined feature amount by learning. The learning application unit 42 may be, for example, a deep learning system. In particular, since the recognition accuracy of image recognition is high, a convolutional neural network (CNN) is preferable for the learning application unit 42.

As an example, the learning application unit 42 composed of this CNN includes, for example, a pretreatment unit 421 and a multi-layer perceptron unit 422, as shown in FIG.

The pretreatment section 421 includes a pretreatment set 4211 (4211-1, 42111-k) including a convolutional layer 42111 (4211-1, ..., 42111-k) and a pooling layer 42112 (4212-1, ..., 42112-k). ..., 4211-k) is configured with one or more k (k is an integer of 2 or more). When there are a plurality of pretreatment sets 4211, these plurality of pretreatment sets 4211 are connected in series to form a plurality of stages. A teacher set for learning and a target sample to be analyzed (recognition, classification, discrimination) are input to the preprocessing unit 421. Here, the plurality of samples in the teacher data and the target sample are images, and hereinafter, the sample of the teacher data and the target sample input to the preprocessing unit 421 are conveniently referred to as an input image IP.

The convolution layer 42111 is a device that executes a convolution operation on the input image IP. More specifically, the convolution layer 42111 convolves (filters) the input image IP with a predetermined kernel (image filter) KN. The kernel KNs are a plurality of preset kernels, and the input image IP is convoluted by each of the plurality of kernel KNs, and a plurality of image (feature map) FMs corresponding to the number of kernel KNs are generated. The more kernel KNs there are, the more various features of the input image IP can be captured. The size of the kernel KN is arbitrary and is appropriately set for each preprocessing set 4211. In CNN, the kernel KN is generated by learning. That is, the numerical value of each element (each filter coefficient) of the kernel KN is automatically set by learning. As a result, the kernel KN acquires a function of extracting a predetermined feature amount by learning. The kernel KN corresponds to an example of the feature amount extraction unit. This convolution operation makes it possible to detect the pattern in the input image IP. The convolution layer 42111 is connected to the pooling layer 42112 in the preprocessing set 4211 of the stage, and outputs the feature map FM, which is an image generated by the convolution operation, to the pooling layer 42112.

The pooling layer 42112 is a device that performs a max pooling process. More specifically, the pooling layer 42112 takes out the maximum value in the rectangle while shifting the rectangular filter of a predetermined size by a predetermined pixel in the feature map FM, and generates a new image. This max pooling process reduces the positional sensitivity of the extracted features. The pooling layer 42112 is connected to the convolution layer 42111 in the pretreatment set 4211 of the next stage when the pretreatment set 4211 of the next stage is present, and the generated image is output to the convolution layer 42111 of the next stage. When the pretreatment set 4211 does not exist (that is, in the case of the pooling layer 42112-k in the final stage pretreatment set 4211-k), it is connected to the multi-layer perceptron section 422 and the generated image is transferred to the multi-layer perceptron section 422. Output.

The multi-layer perceptron unit 422 is a neural network in which neurons are arranged in a plurality of layers, and is generated by the neural network in the processing result of the preprocessing unit 421 (the pooling layer 42112-k in the preprocessing set 4211-k of the final stage). This is a device that recognizes the input image IP input to the preprocessing unit 421 based on the image). The multi-layer perceptron unit 422 outputs the OR as a result of the recognition.

As an example, such a learning model creating device Da includes a CPU 11, a RAM 12, a ROM 13, a hard disk drive (HDD) 14, an LCD 16, a keyboard 17, a mouse 18, a USB interface 19, and these CPUs 11. It can be configured by, for example, a desktop or node-type computer including a bus 15 that interconnects a RAM 12, a ROM 13, an HDD 14, an LCD 16, a keyboard 17, a mouse 18, and a USB interface 19.

Next, the operation of this embodiment will be described. FIG. 4 is a flowchart showing the operation of the learning model creating device in learning. FIG. 5 is a flowchart showing the operation of the learning model creating device in the analysis. FIG. 6 is a flowchart showing a redundant deletion process in each of the flowcharts shown in FIGS. 4 and 5. FIG. 7 is a diagram showing the relationship between the number of repetitions of machine learning and the total amount of information in the feature amount extraction unit. The horizontal axis of FIG. 7 is the number of iterations (number of iterations) iter of machine learning, and the vertical axis thereof is the total amount of information of the feature amount extraction unit in the learning model. FIG. 8 is for explaining the redundant deletion process based on the contribution in the first embodiment. FIG. 9 is a diagram for explaining the effect of the learning model creating device in the first embodiment.

When the power of the learning model creation device Da having such a configuration is turned on, the necessary initialization of each part is executed and the operation is started. By executing the control processing program, the control processing unit 4a is functionally configured with a control unit 41, a learning application unit 42, a specific unit 43a, and a model improvement unit 44.

Then, in the learning process for generating the learning model, in FIG. 4, in FIG. 4, in the learning model creating device Da, the learning in which the processes S11, S12, and S13 from the process # 1 to the process # 2 are preset by the control process unit 4a. The process S21a and the process S22 are executed at the improvement timing determined by the process S13 after being repeated by the number of times (the number of repetitions).

More specifically, in the process S11, the learning model creating device Da determines whether or not the control variable i is larger than the number of repetitions max_itter by the learning application unit 42 of the control processing unit 4a. The control variable i is a variable representing the number of repetitions of each process from process # 1 to process # 2. The repetition number max_itter is the number of repetitions (repetition number) of learning using the teacher set preset by the user (operator), and is appropriately set at, for example, 100 epochs, 500 epochs, 1000 epochs, and the like. As a result of this determination, when the control variable i is not larger than the number of repetitions max_ita (No), learning is in progress, and the learning application unit 42 next executes the process S12. On the other hand, as a result of the determination, when the control variable i is larger than the number of repetitions max_ita (Yes), the learning is completed, and the learning application unit 42 repeats each process from the process # 1 to the process # 2. Then, the process S14 is executed.

In the process S12, the learning model creation device Da executes machine learning by the learning application unit 42 using the teacher set (teacher data and correct answer label) stored in the teacher set storage unit 51 of the storage unit 5, and then performs machine learning. , Process S13 is executed.

In the process S13, the learning model creation device Da deletes the redundant feature amount extraction unit from the learning model being trained generated in the process S12 by the control processing unit 4a, and whether or not it is an improvement timing for improving the learning model. Is determined. If the process of deleting the redundant feature amount extraction unit is always performed for each epoch, the performance of the learning model may deteriorate and the calculation cost increases. Therefore, the improvement timing is adjusted in such process S13. By determining, these can be avoided.

This improvement timing is appropriately determined by the user (operator) according to, for example, the processing capacity of the hardware constituting the learning model creation device Da, the type of analysis processing, the expected progress of learning, and the like, and the learning model is created. It is input and set in advance in the device Da. Alternatively, the difference between the learning models is obtained before and after one epoch, and when this difference is equal to or less than a preset predetermined threshold value, it is automatically determined that the improvement timing has arrived. For example, when a learning model is generated by learning 100 epochs using predetermined teacher data, if the processing power of the hardware is high for the type of analysis processing, the first epoch is set as the improvement timing, and the learning model. May be improved, and for example, in such a case, the improvement timing may be set every few epochs, and the learning model may be improved. Further, for example, if the predicted progress of learning is slow, the learning model may be improved by avoiding the first few epochs and setting the improvement timing every 10 epochs, for example. When the improvement timing is set in multiple stages in this way, as shown in FIG. 7, in the conventional general machine learning, the total amount of information in the feature amount extraction unit is the total amount of information regardless of the number of repetitions (number of repetitions) itter. , B [Byte] is constant, but in the present embodiment, the total amount of information of the feature amount extraction unit at the time of an arbitrary number of repetitions ix is α × b [when the number of deletion settings is α. Byte] is reduced. As a result, the calculation cost required for iterative learning is gradually reduced. Furthermore, in machine learning, the learning model is improved so that the correct answer can be derived using the past learning model and its performance, so it is necessary to refer to it for improvement by deleting the feature quantity extraction unit. Since there is no more, the learning model can converge faster than before. In such a multi-stage improvement timing, when the required accuracy for actual application in the learning model is obtained by the number of iterations Re, the required accuracy for actual application is taken into consideration so that ix ≦ Re. The number of iterations ix is set to.

As a result of the determination of the process S13, if it is not the improvement timing (No), the learning of the epoch is completed and the learning of the next epoch is started. The control variable i is incremented by 1 (i = i + 1), and the process is returned to the process S11.

On the other hand, as a result of the determination of the process S13, when the improvement timing is reached, the learning model creation device Da increments the control variable i by 1 after each of the process S21a and the process S22 is executed by the control process unit 4a. (I = i + 1), the process is returned to the process S11. In the process S21a, in order to delete the redundant feature amount extraction unit, a deletion target specifying process for specifying the feature amount extraction unit to be deleted is executed, and in the process S22, the feature amount extraction unit specified in the process S21a is executed. A part of the feature amount extraction unit that deletes from the learning model and generates a new learning model is deleted. Therefore, in the next epoch, learning is performed by a new learning model in which the redundant feature amount extraction unit is deleted from the learning model obtained by the machine learning of the process S12. Each of these processes S21a and S22 will be described in detail later.

Then, when machine learning of the number of repetitions max_itter is executed, the process S14 is executed as described above, and in this process S14, the learning model creating device Da outputs the learning model created by the control processing unit 4a. Then, this process is terminated.

On the other hand, in the analysis process for analyzing the input sample, in FIG. 5, the learning model creating device Da executes the specific process S21a to be deleted by the control process unit 4a, and then a part of the feature amount extraction unit. Deletion process S22 is executed. As a result, a new learning model is created from the original learning model (for example, a learning model after machine learning) by removing the redundant feature amount extraction unit. Then, the learning model creating device Da analyzes the target sample with the new learning model created in the processing S22 by the learning application unit 42 of the control processing unit 4a (S31), and outputs the analysis result (S32). , End this process. Therefore, the analysis of the target sample is performed with a new learning model in which the redundant feature amount extraction unit is deleted from the initial learning model.

Next, the specific processing S21a to be deleted and the deletion processing S22 of a part of the feature amount extraction unit will be described more specifically.

In FIG. 6, the learning model creation device Da acquires a specific condition by the control processing unit 4a (S41). As an example, the specific condition is the number of feature quantity extraction units to be deleted as the redundant feature quantity extraction unit (deletion set number) in the present embodiment. The number of deletion settings is input and acquired from the input unit 1, for example, numerically or at a ratio to the total number of feature quantity extraction units. Since there is a trade-off relationship between the amount of reduction in calculation cost and the deterioration of performance for the learning model generated by deleting the redundant feature quantity extraction unit, the number of deletion settings can be set in the process S41. By configuring the learning model creation device Da, the user can set the number of deletion settings in consideration of this trade-off relationship.

Subsequently, in the learning model creation device Da, each process S42, S43, S44 from process # 111 to process # 112 is repeated by the control process unit 4a by the number of feature amount extraction units, and then the process S45, process S46, and processing S46 and Each process of process S47 is sequentially performed.

More specifically, in the process S42, the learning model creating device Da deletes only the kth feature amount extraction unit from the learning model by the specific unit 43a of the control processing unit 4a. As a result, a new learning model K for trial is generated ((learning model K for trial) = (learning model in which only the k-th feature amount extraction unit is deleted from the learning model)). In the case of the learning process, the learning model is the learning model generated in the above-mentioned process S12 shown in FIG. 4, and in the case of the analysis process, it is the initial learning model.

In the process S43 following the process S42, the learning model creation device Da analyzes (determines, classifies) the teacher data with the trial learning model K generated in the process S42 by the specific unit 43a of the control process unit 4a. implement.

In the process S44 following the process S43, the learning model creation device Da obtains the contribution degree in the learning model K for trial by the specific unit 43a of the control processing unit 4a, increments k by 1 (k = k + 1), and processes. Is returned to the process S42. More specifically, the specific unit 43a evaluates the performance of the learning model K for trial by matching the analysis result (classification result) obtained in the process S43 with the correct answer label corresponding to the teacher data. The evaluation value to be evaluated is obtained as the degree of contribution in the k-th feature amount extraction unit. Before and after the deletion of only the k-th feature amount extraction unit, it is determined that the smaller the deterioration of the evaluation value, the smaller the contribution degree, and the feature amount extraction unit having a small contribution degree is a redundant feature amount extraction unit. can. The evaluation value is appropriately defined in advance by the user depending on, for example, the type of analysis processing. For the evaluation value AC, for example, when the correct answer label is represented by the binary classification that classifies one sample into one of the binary values, the total number of samples of the teacher data is N, and the correct answer sample that can be correctly classified. When the number is C, it is represented by C / N (AC = C / N). Further, for example, when the correct answer label is represented by the multi-class classification that classifies one sample into multiple classes, the evaluation value AC is the correct sample that can be correctly classified because the total number of samples of the teacher data is N. When the number is C, it is represented by C / N (AC = C / N). In multi-class classification, the probability of each class is output for one sample, and if the class with the highest probability is the correct answer, it is only the case that the sample can be correctly discriminated (analyzed and classified). If, for example, a class having a probability within the top 3 is correct, it may be considered that the sample can be correctly discriminated (analyzed and classified), depending on the type of analysis processing. Further, for example, in the regression problem, the evaluation value AC is given as an error value which is an amount of deviation from the correct answer.

Then, when each of the processes S42 to S44 is performed for each of the plurality of feature amount extraction units in the learning model to obtain a plurality of contributions in each of the plurality of feature amount extraction units, the subsequent process S45 The learning model creating device Da arranges (sorts) the plurality of feature quantity extraction units in order from the one with the lowest contribution by the specific unit 43a of the control processing unit 4a. As a result, a contribution list for the plurality of feature quantity extraction units is generated. An example is shown in FIG. In FIG. 8, the horizontal axis is the degree of contribution, and each ● mark represents the feature amount extraction unit.

In the process S46 following the process S45, the learning model creation device Da performs a specific process of specifying the feature amount extraction unit to be deleted by the specific unit 43a of the control process unit 4a. It can be estimated that the feature amount extraction unit having a small contribution has a small influence on the performance of the learning model and is a redundant feature amount extraction unit. On the other hand, the feature amount extraction unit having a large contribution (for example, the feature amount extraction unit in the boxed line in FIG. 8) is presumed to be an important feature amount extraction unit that has a large influence on the performance of the learning model and is not redundant. can. Therefore, in the present embodiment, the specific unit 43a selects the feature amount extraction unit in order from the lowest to the deletion set number, and specifies the selected feature amount extraction unit as the redundant feature amount extraction unit. .. That is, the specifying unit 43a identifies the redundant feature amount extracting unit in order from the feature amount extracting unit having the lowest contribution while satisfying the specific conditions accepted by the input unit. When the user inputs x% of the total number of feature quantity extraction units as the deletion setting number from the input unit 1, the feature quantity extraction unit from the lowest to the one-dot chain line shown in FIG. 8 is selected, and the redundant feature is selected. Specified as a quantity extractor. As a result, it is possible to identify a feature amount extraction unit that can be deleted while maintaining the performance of the learning model.

In the process S47 following the process S46, the learning model creation device Da is new by the model improvement unit 44 of the control processing unit 4a deleting the feature amount extraction unit specified by the specific unit 43a in the process S46 from the learning model. Create a learning model and end this process. As a result, a new learning model whose size is as small as the number of deleted feature quantity extraction units can be created.

When the learning application unit 42 is a CNN having a configuration of a plurality of stages, each of the above-mentioned processes S41 to S47 may be performed on any stage (layer), and one stage may be used. It may be implemented for (layers), or may be implemented for a plurality of stages (layers), and may be appropriately set for the user (operator). From the viewpoint of effectively reducing the calculation cost, it is preferable that each of the above-mentioned processes S41 to S47 is performed on the first stage (first layer).

As described above, the learning model creation device Da in the first embodiment, and the learning model creation method and learning model creation program implemented therein use the learning model during or after learning. When analyzing the input sample, a redundant feature amount extraction unit is specified from a plurality of feature amount extraction units, and the specified feature amount extraction unit is deleted, so that the performance deterioration of the learning model can be suppressed. Then, the learning model creation device Da, the learning model creation method, and the learning model creation program proceed with learning with a new learning model of a smaller size by deleting the feature amount extraction unit during learning, so that the calculation cost of learning is reduced. Can be reduced. In addition, the learning model creation device Da, the learning model creation method, and the learning model creation program analyze the input sample with a new learning model of a smaller size by deleting the feature amount extraction unit at the time of analysis. The calculation cost of can be reduced. Therefore, the learning model creation device Da, the learning model creation method, and the learning model creation program can reduce the calculation cost while suppressing the performance deterioration of the learning model.

FIG. 9 is a diagram for explaining the effect of the learning model creating device in the first embodiment. In one specific example, as shown in FIG. 9, when a CNN learning model is created by the conventional method using a plurality of images of trees such as cedars and fir trees as teacher data, during learning or learning. The later learning model includes a feature amount extraction unit A that reacts to green, a feature amount extraction unit B that reacts to an upward-sloping edge, a feature amount extraction unit C that reacts to an upward-sloping edge, and a feature amount extraction unit that reacts to green. A plurality of feature amount extraction units including M are included. Therefore, when a learning model is created using the teacher data by the learning model creation device Da in the first embodiment, one of the feature amount extraction unit A and the feature amount extraction unit M is a redundant feature amount extraction unit. , One feature quantity extraction unit is deleted, thereby creating a learning model having a smaller size, and the other feature quantity extraction unit remains, so that the performance deterioration of the learning model is suppressed.

In another specific example, the experiment was carried out in the following environment. In this experimental environment, the framework is "caffe", the dataset is "The CIFAR-10 dataset", the input image name is "canis_familiaris_s_001129.png", and the network definition is "AlexNet". In such an experimental environment, when an original learning model in which the feature amount extraction unit is not deleted is used (when the number of deletion settings is 0), that is, the learning model creation device Da of the present embodiment is used for improvement. In the case of using a learning model that has not been used, when the correct answer class is the dog class, 58.83% of the classification results are classified as dog, 23.91% are classified as cat, and 7.92% are classified as horse. It was classified into. On the other hand, in this original learning model, the learning model creation device Da in the present embodiment sets the number of deletion settings to 50% of the total number of feature quantity extraction units, and extracts the redundant feature quantity with respect to the first layer of AlexNet. Improved to remove the part. The feature amount extraction unit (kernel) in the first layer of AlexNet has a size of 11 × 11, and the total number is 96. When the learning model improved in this way is used, when the correct answer class is the dog class, 63.45% of the classification results are classified as dog and 26.72% are classified as cat. 63% were classified as horses. As can be seen by comparing these, even if the feature amount extraction unit is deleted by 50%, it can be correctly classified into the correct dog class, and the learning model creation device Da in the present embodiment suppresses the performance deterioration of the learning model. At the same time, by deleting the feature amount extraction unit, a new learning model with a smaller size can be created, and the calculation cost can be reduced.

Since the learning model creation device Da, the learning model creation method, and the learning model creation program include an input unit 1 that accepts input of the number of deletion settings, the user can reduce the calculation cost and deteriorate the performance, which are in a trade-off relationship. The number of deletion settings can be set in consideration of.

Since the learning model creation device Da, the learning model creation method, and the learning model creation program include an input unit 1 that accepts input of improvement timing, the user can use, for example, the processing capacity of the hardware constituting the learning model creation device Da. The timing can be set in consideration of the type of analysis processing, the expected progress of learning, and the like.

The learning model creation device Da, the learning model creation method, and the learning model creation program can appropriately specify the redundant feature amount extraction unit from the viewpoint of contribution. In the learning model creation device Da, the learning model creation method, and the learning model creation program, since the learning model is specified from the viewpoint of the degree of contribution, the redundant feature amount extraction unit is specified with priority given to the reduction of the calculation cost.

Next, another embodiment will be described.

(Second Embodiment)
FIG. 10 is for explaining the redundant deletion process by the sum of squared differences in the second embodiment.

The learning model creating device Da in the first embodiment specified the redundant feature amount extracting unit in order from the feature amount extracting unit having the lowest contribution, but the learning model creating device Db in the second embodiment is in the learning model. The redundant feature amount extraction unit is specified in order from the feature amount extraction unit having low independence among the plurality of feature amount extraction units.

As shown in FIG. 1, for example, the learning model creating device Db in the second embodiment includes an input unit 1, an output unit 2, an IF unit 3, a control processing unit 4b, and a storage unit 5. Be prepared. The input unit 1, the output unit 2, the IF unit 3 and the storage unit 5 in the learning model creation device Db in the second embodiment are the input unit 1, the output unit 2 and the output unit 2 in the learning model creation device Da in the first embodiment, respectively. Since it is the same as the IF unit 3 and the storage unit 5, the description thereof will be omitted.

The control processing unit 4b is a circuit for creating a learning model for machine learning by controlling each part 1 to 3 and 5 of the learning model creating device Db according to the function of each part, and is, for example, a CPU and its surroundings. It is configured with a circuit. The control processing unit 4b functionally includes a control unit 41, a learning application unit 42, a specific unit 43b, and a model improvement unit 44 by executing the control processing program. The control unit 41, the learning application unit 42, and the model improvement unit 44 in the control processing unit 4b of the second embodiment are the control unit 41, the learning application unit 42, and the model improvement unit in the control processing unit 4a of the first embodiment, respectively. Since it is the same as 44, the description thereof will be omitted.

The specific unit 43b identifies a redundant feature amount extraction unit from a plurality of feature amount extraction units in the learning model during training or when analyzing an input sample using the learning model after learning. It is something to do. In the present embodiment, the specific unit 43b specifies the redundant feature amount extraction unit with priority given to reduction of calculation cost. More specifically, the specific unit 43b identifies the redundant feature amount extraction unit in order from the feature amount extraction unit having a low degree of independence indicating the degree of independence between the feature amount extraction units in the learning model. More specifically, the specific unit 43b obtains the difference between the components between the feature amount extraction units, obtains the sum of the absolute values of the obtained differences as the degree of independence, and based on the obtained degree of independence, the above-mentioned The redundant feature amount extraction unit is specified in order from the feature amount extraction unit having the lowest degree of independence. More specifically, in the specific unit 43b, first, for each of the plurality of feature amount extraction units, the feature amount extraction unit and the residual feature amount extraction unit obtained by removing the feature amount extraction unit from the plurality of feature amount extraction units, respectively. Find the difference between the components with. Next, the specific unit 43b obtains the sum of the squares of the obtained differences (the value obtained by accumulating all the squares of the differences) as the degree of independence (independence = sum of the squares of the differences (cumulative value of the squares of the difference)). The smaller the sum of squared differences, the smaller the degree of independence, and it can be determined that the feature amount extraction unit having a small degree of independence is a redundant feature amount extraction unit. An example is shown in FIG. The horizontal axis of FIG. 10 shows the number (kernel_num) of the kernel KN which is an example of the feature amount extraction unit, and the vertical axis thereof shows the sum of squares (sum of squares), that is, the degree of independence. Then, the specific unit 43b arranges the plurality of feature quantity extraction units in order from the one with the smallest degree of independence (difference squared sum), and the feature quantity extraction units are sequentially arranged from the lowest to the preset deletion set number. It is selected, and the selected feature amount extraction unit is specified as the redundant feature amount extraction unit.

The learning model creating device Db in such a second embodiment is shown in FIG. Each of the above-mentioned processes is carried out using 4. As a result, a learning model in which the redundant feature amount extraction unit is deleted is created.

On the other hand, in the analysis process, the learning model creation device Db in the second embodiment performs the deletion target specific process S21b by the specific unit 43b instead of the deletion target specific process S21a by the specific unit 43a. Each of the above-mentioned treatments is carried out using. As a result, the analysis of the target sample is performed with a new learning model in which the redundant feature amount extraction unit is removed from the initial learning model.

In the specific processing S21b to be deleted, the learning model creating device Db obtains the difference between the components between the feature amount extraction units by the specific unit 43b, and the sum of the absolute values of the obtained differences is the degree of independence. Based on this determined degree of independence, from the feature amount extraction unit having the lowest independence among the plurality of feature amount extraction units, until the number of deleted features is reached (that is, in the input unit). (While the accepted specific conditions are satisfied), the redundant feature amount extraction unit is specified.

The learning model creation device Db in the second embodiment, and the learning model creation method and the learning model creation program implemented therein are the learning model creation device Da, the learning model creation method, and the learning model in the first embodiment. It has the same effect as the created program.

Then, the learning model creation device Db, the learning model creation method, and the learning model creation program can appropriately specify the redundant feature amount extraction unit from the viewpoint of the degree of independence. Since the learning model creation device Db, the learning model creation method, and the learning model creation program are specified from the viewpoint of the degree of independence, the redundant feature amount extraction unit is specified with priority given to reduction of calculation cost.

Next, another embodiment will be described.

(Third Embodiment)
FIG. 11 is for explaining the redundant deletion process by dimensional compression in the third embodiment. FIG. 11A shows a learning model when the function of the feature quantity extraction unit is represented by text, FIG. 11B shows a learning model when the components of the feature quantity extraction unit are represented by a vector, and FIG. 11C is two-dimensional. The distribution of each feature amount extraction part when compressed to is shown.

The learning model creating device Da in the first embodiment specified the redundant feature amount extracting unit in order from the feature amount extracting unit having the lowest contribution, but the learning model creating device Dc in the third embodiment is in the learning model. A plurality of feature amount extraction units are grouped, and the remaining feature amount extraction unit excluding the feature amount extraction unit representing each group from the plurality of feature amount extraction units is specified as the redundant feature amount extraction unit. Is.

As shown in FIG. 1, for example, the learning model creating device Dc in the third embodiment includes an input unit 1, an output unit 2, an IF unit 3, a control processing unit 4c, and a storage unit 5. Be prepared. The input unit 1, the output unit 2, the IF unit 3 and the storage unit 5 in the learning model creation device Dc in the third embodiment are the input unit 1, the output unit 2 and the output unit 2 in the learning model creation device Da in the first embodiment, respectively. Since it is the same as the IF unit 3 and the storage unit 5, the description thereof will be omitted.

The control processing unit 4c is a circuit for controlling each unit 1 to 3 and 5 of the learning model creation device Dc according to the function of each unit to create a learning model for machine learning. For example, the CPU and its surroundings. It is configured with a circuit. The control processing unit 4c functionally includes a control unit 41, a learning application unit 42, a specific unit 43c, and a model improvement unit 44 by executing the control processing program. The control unit 41, the learning application unit 42, and the model improvement unit 44 in the control processing unit 4c of the third embodiment are the control unit 41, the learning application unit 42, and the model improvement unit in the control processing unit 4a of the first embodiment, respectively. Since it is the same as 44, the description thereof will be omitted.

The specific unit 43c identifies a redundant feature amount extraction unit from a plurality of feature amount extraction units in the learning model during learning or when analyzing an input sample using the learning model after learning. It is something to do. In the present embodiment, the specific unit 43c specifies the redundant feature amount extraction unit with priority given to reduction of calculation cost. More specifically, the specific unit 43c divides a plurality of feature quantity extraction units in the learning model into a plurality of groups including feature quantity extraction units similar to each other, and features representing the group from each of the plurality of groups. By selecting the amount extraction unit as the representative feature amount extraction unit and specifying the residual feature amount extraction unit excluding the representative feature amount extraction unit from the plurality of feature amount extraction units as the redundant feature amount extraction unit, the above-mentioned Identify redundant feature extraction units. More specifically, the specific unit 43c first obtains a distance between the plurality of feature amount extraction units in order to group the plurality of feature amount extraction units. It can be determined that the feature amount extraction units that are close to each other are feature amount extraction units that are similar to each other. The distance may be obtained in a dimension corresponding to the number of components of the feature amount extraction unit, but in the present embodiment, it is obtained after reducing the dimension of the feature amount extraction unit. For example, the specific unit 43c reduces the dimension of the plurality of feature quantity extraction units by using a method of principal component analysis. In one example, when the feature quantity extraction unit of the learning model shown in FIG. 11A consists of n-dimensional components (n components) as shown in FIG. 11B, the first and second principal components are analyzed using the method of principal component analysis. As shown in FIG. 11C, a plurality of feature quantity extraction units compressed in two dimensions of the components and compressed in two dimensions are required. Next, the specific unit 43c obtains a distance between the reduced-dimensional feature amount extraction units, and groups the feature amount extraction units whose distance is within a predetermined threshold value (distance determination threshold value) set in advance. In the example shown in FIG. 11C, the plurality of feature amount extraction units are divided into seven first to seventh groups GPa to GPg. Next, the specific unit 43c selects a feature amount extraction unit representing the group as the representative feature amount extraction unit from each of the plurality of groups. In the example shown in FIG. 11C, the specific unit 43c obtains the position of the center of gravity of the group for each group GP, and selects the feature amount extraction unit closest to the obtained center of gravity position in terms of distance as the representative feature amount extraction unit. do. In the example shown in FIG. 11C, in the first group GPa, the feature amount extraction unit A is selected as the representative feature amount extraction unit, and in the second group GPb, the feature amount extraction unit B is selected as the representative feature amount extraction unit. In the 3 group GPc, the feature amount extraction unit C is selected as the representative feature amount extraction unit, in the 4th group GPd, the feature amount extraction unit D is selected as the representative feature amount extraction unit, and in the 5th group GPe, the feature amount extraction unit is selected. Part E is selected as the representative feature amount extraction unit, and in the sixth group GPf, the feature amount extraction unit F is selected as the representative feature amount extraction unit. Next, the specific unit 43c specifies the residual feature amount extraction unit excluding the representative feature amount extraction unit from the plurality of feature amount extraction units as the redundant feature amount extraction unit.

In the third embodiment, when the number of deletion settings is set, the distance determination threshold value is such that the total number of groups after grouping is the total number of feature amount extraction units minus the number of deletion settings. Is adjusted. As a result, the specific unit 43c is the number obtained by subtracting the number of deletion settings received by the input unit 1 from the total number of the plurality of feature amount extraction units in the plurality of groups of the plurality of feature amount extraction units in the learning model. So, divide.

The learning model creating device Dc in such a third embodiment is shown in FIG. Each of the above-mentioned processes is carried out using 4. As a result, a learning model in which the redundant feature amount extraction unit is deleted is created.

On the other hand, in the analysis process, the learning model creation device Dc in the third embodiment performs the deletion target specific process S21c by the specific unit 43c instead of the deletion target specific process S21a by the specific unit 43a. Each of the above-mentioned treatments is carried out using. As a result, the analysis of the target sample is performed with a new learning model in which the redundant feature amount extraction unit is removed from the initial learning model.

In the specific processing S21c to be deleted, the learning model creation device Dc divides a plurality of feature amount extraction units in the learning model into a plurality of groups including feature amount extraction units similar to each other by the specific unit 43c, and the plurality of features are extracted. From each of the groups, the feature amount extraction unit representing the group is selected as the representative feature amount extraction unit, and the residual feature amount extraction unit excluding the representative feature amount extraction unit from the plurality of feature amount extraction units is the redundant feature. Specify as a quantity extractor.

The learning model creation device Dc in the third embodiment, and the learning model creation method and the learning model creation program implemented therein are the learning model creation device Da, the learning model creation method, and the learning model in the first embodiment. It has the same effect as the created program.

Then, the learning model creation device Dc, the learning model creation method, and the learning model creation program can appropriately identify the redundant feature amount extraction unit from a similar viewpoint. Since the learning model creation device Dc, the learning model creation method, and the learning model creation program are specified from the same viewpoint, the redundant feature amount extraction unit is specified with priority given to reduction of calculation cost.

Next, another embodiment will be described.

(Fourth Embodiment)
FIG. 12 is a diagram showing an example of an input / output screen displayed on the learning model creating device according to the fourth embodiment.

The learning model creating devices Da to Dc in the first to third embodiments specify the redundant feature amount extraction unit with priority given to reduction of calculation cost, but the learning model creating device Dd in the fourth embodiment is used for learning. The redundant feature amount extraction unit is specified with priority given to maintaining the performance of the model.

As shown in FIG. 1, for example, the learning model creating device Dd in the fourth embodiment includes an input unit 1, an output unit 2, an IF unit 3, a control processing unit 4d, and a storage unit 5. Be prepared. In the input unit 1, the output unit 2, the IF unit 3, and the storage unit 5 in the learning model creation device Dd in the fourth embodiment, the input unit 1 further sets a first priority item for prioritizing the maintenance of the performance of the learning model. Except for the point of acceptance, it is the same as the input unit 1, the output unit 2, the IF unit 3 and the storage unit 5 in the learning model creation device Da in the first embodiment, and thus the description thereof will be omitted.

In the present embodiment, the input unit 1 further accepts the input of the first priority item that prioritizes the maintenance of the performance of the learning model, as described above. In the present embodiment, the first priority item is a first feature amount extraction unit (color information extraction unit) that extracts color information as the feature amount, and a second feature amount extraction unit that extracts edge information as the feature amount (the first feature amount extraction unit). Features selected from the edge information extraction unit) and the third feature amount extraction unit (blurring information extraction unit) that extracts blur information as the feature amount, and excluded from the redundant feature amount extraction unit to be specified. It is a quantity extraction unit. The selected feature amount extraction unit may be one, may be plural, or may be all. The first priority item is another example of the specific condition. Therefore, the input unit 1 receives the input of the feature amount extraction unit excluded from the redundant feature amount extraction unit to be specified from the first to third feature amount extraction units.

The control processing unit 4d is a circuit for creating a learning model for machine learning by controlling each unit 1 to 3 and 5 of the learning model creating device Dd according to the function of each unit. For example, the CPU and its surroundings. It is configured with a circuit. The control processing unit 4d functionally includes a control unit 41, a learning application unit 42, a specific unit 43d, and a model improvement unit 44 by executing the control processing program. The control unit 41, the learning application unit 42, and the model improvement unit 44 in the control processing unit 4d of the fourth embodiment are the control unit 41, the learning application unit 42, and the model improvement unit in the control processing unit 4a of the first embodiment, respectively. Since it is the same as 44, the description thereof will be omitted.

The specific unit 43d identifies a redundant feature amount extraction unit from a plurality of feature amount extraction units in the learning model during learning or when analyzing an input sample using the learning model after learning. It is something to do. In the present embodiment, the specific unit 43d specifies the redundant feature amount extraction unit with priority given to maintaining the performance of the learning model. More specifically, the specific unit 43d excludes the feature amount extraction unit corresponding to the feature amount extraction unit received by the input unit 1 from the redundant feature amount extraction unit to be specified, whereby the redundant feature is described. Identify the quantity extractor. That is, when the feature amount extraction unit specified as the redundant feature amount extraction unit is the feature amount extraction unit corresponding to the feature amount extraction unit received by the input unit 1, the specific unit 43d is said to be redundant. Not specified as a feature amount extraction unit. More specifically, for example, when the redundant feature amount extraction unit is specified in order from the feature amount extraction unit having the lowest contribution, the feature amount extraction unit corresponding to the feature amount extraction unit received by the input unit 1 contributes to this. It is removed from the plurality of feature amount extraction units arranged in order of degree, and the specific unit 43d removes the feature amount extraction unit corresponding to the feature amount extraction unit received by the input unit 1 and performs the redundancy in order from the feature amount extraction unit having the lowest contribution. Identify a feature amount extraction unit. Further, for example, when the redundant feature amount extraction unit is specified in order from the feature amount extraction unit having the lower degree of independence, the feature amount extraction unit corresponding to the feature amount extraction unit received by the input unit 1 is in the order of the independence degree. The specific unit 43d is removed from the plurality of feature quantity extraction units arranged side by side, and the feature quantity extraction unit corresponding to the feature quantity extraction unit received by the input unit 1 is removed, and the redundant features are sequentially described from the feature quantity extraction unit having the lowest degree of independence. Specify the quantity extraction unit. Further, for example, a plurality of feature amount extraction units in the learning model are divided into a plurality of groups consisting of feature amount extraction units similar to each other, and residual feature amount extraction excluding the representative feature amount extraction unit is extracted from each of the plurality of groups. When the unit is specified as the redundant feature amount extraction unit, the feature amount extraction unit corresponding to the feature amount extraction unit received by the input unit 1 is removed from the residual feature amount extraction unit, and the specific unit 43d is input. The feature amount extraction unit corresponding to the feature amount extraction unit received in Part 1 is removed, and the residual feature amount extraction unit is specified.

The learning model creating device Dd in the fourth embodiment is shown in the figure except that in the learning process, the deletion target specific process S21d by the specific unit 43d is performed instead of the deletion target specific process S21a by the specific unit 43a. Each of the above-mentioned processes is carried out using 4. As a result, a learning model in which the redundant feature amount extraction unit is deleted is created.

On the other hand, in the analysis process, the learning model creating device Dd in the fourth embodiment performs the deletion target specific process S21d by the specific unit 43d instead of the deletion target specific process S21a by the specific unit 43a. Each of the above-mentioned treatments is carried out using. As a result, the analysis of the target sample is performed with a new learning model in which the redundant feature amount extraction unit is removed from the initial learning model.

In the deletion target specific processing S21d, the learning model creation device Dd receives the first priority item in the input unit 1 by the specific unit 43d, and the feature amount extraction unit of the first priority item received by the input unit 1 is used. The redundant feature amount extraction unit is specified by excluding the corresponding feature amount extraction unit from the redundant feature amount extraction unit to be specified.

The input of the first priority item is performed using, for example, the input / output screen shown in FIG. The input / output screen 7 shown in FIG. 12 inputs the deletion setting number input area 71 for inputting the deletion setting number, which is an example of the specific condition, and the first priority item, which is another example of the specific condition. The priority item input area 72 for displaying the original learning model before the redundant feature amount extraction unit is deleted, the original learning model display area 73 for displaying the original learning model in the feature amount extraction unit, and the redundant feature amount extraction unit. A post-deletion learning model display area 74 for displaying the learning model after the section is deleted by the feature amount extraction section is provided.

In the example shown in FIG. 12, the deletion setting number area 71 is configured so that the deletion setting number can be input as a ratio to the total number of feature amount extraction units. More specifically, the deletion setting number area 71 includes a scale bar 711 representing a ratio from 0% to 100%, and a ratio designation cursor 712 that specifies the ratio by changing the position on the scale bar 711. , A designated ratio numerical display area 713 for displaying the ratio designated by the ratio cursor as a numerical value is provided. In the example shown in FIG. 12, the number of deletion settings is the number corresponding to 20% of the total number of feature amount extraction units by the ratio designation cursor 712.

In the example shown in FIG. 12, the priority item input area 72 is for designating whether to exclude or not exclude the third feature amount extraction unit that extracts blur information as the feature amount from the feature amount extraction unit to be specified. A first check box 721, a second check box 722 for designating whether to exclude or not to exclude the first feature amount extraction unit that extracts color information as the feature amount from the specified feature amount extraction unit, and the above. It is provided with a third check box 723 for designating whether to exclude or not exclude the second feature amount extraction unit that extracts edge information as the feature amount from the feature amount extraction unit to be specified. When the first check box 721 is checked, the third feature amount extraction unit that extracts the blur information as the feature amount is excluded from the specified feature amount extraction unit, and therefore the blur information is not erased (deleted). When the second check box 722 is checked, the first feature amount extraction unit that extracts color information as the feature amount is excluded from the specified feature amount extraction unit, and therefore the color information is not erased (deleted). When the third check box 723 is checked, the second feature amount extraction unit that extracts edge information as the feature amount is excluded from the specified feature amount extraction unit, and therefore the edge information is not erased (deleted). In the example shown in FIG. 12, only the third check box 723 is checked.

In the example shown in FIG. 12, the original learning model display area 73 displays an original learning model having 96 feature quantity extraction units, and the deleted learning model display area 74 displays 96 feature quantity extraction units. A learning model after deletion is displayed in which 19 of the feature quantity extraction units are deleted as the redundant feature quantity extraction unit. The feature amount extraction unit deleted as the redundant feature amount extraction unit is shown in black in FIG. 12. Since such an input / output screen 7 includes an original learning model display area 73 and a post-deletion learning model display area 74, the user is displayed in each of the original learning model display area 73 and the post-deletion learning model display area 74. While comparing each feature amount extractor, the number of deletion settings can be set in consideration of the trade-off relationship.

The learning model creation devices Da to Dc in the first to third embodiments use the input / output screen 7 not provided with the priority item input area 72, and the number of deletion settings is input to the learning model creation devices Da to Dc. And may be set. Further, on the input / output screen 7, the redundant feature amount extraction unit is specified with priority given to maintaining the performance of the learning model depending on whether or not at least one of the first to third check boxes 721 to 723 is checked. Alternatively, the input of an instruction as to whether to specify the redundant feature amount extraction unit with priority given to the reduction of the calculation cost may be determined.

The learning model creation device Dd in the fourth embodiment, and the learning model creation method and the learning model creation program implemented therein are the learning model creation device Da, the learning model creation method, and the learning model in the first embodiment. It has the same effect as the created program.

The learning model creation device Dd, the learning model creation method, and the learning model creation program can maintain the performance corresponding to the first priority item received by the input unit 1, and can be used for the purpose of the learning model according to the user's intention. The performance of the learning model can be maintained by leaving the corresponding important feature amount extraction unit. When the input unit 1 is accepted to exclude the first feature amount extraction unit (color information extraction unit), the learning model creation device Dd, the learning model creation method, and the learning model creation program are used to extract the first feature amount. Since the redundant feature amount extraction unit is specified by excluding the unit, the first feature amount extraction unit can be left in the learning model after the redundant feature amount extraction unit is deleted. In this case, since the first feature amount extraction unit is left in the learning model creation device Dd, the learning model creation method, and the learning model creation program, the redundant feature amount extraction unit is specified with priority given to maintaining the performance of the learning model. Will be done. When the input unit 1 is accepted to exclude the second feature amount extraction unit (edge information extraction unit), the learning model creation device Dd, the learning model creation method, and the learning model creation program are used to extract the second feature amount. Since the redundant feature amount extraction unit is specified by excluding the unit, the second feature amount extraction unit can be left in the learning model after the redundant feature amount extraction unit is deleted. In this case, since the second feature amount extraction unit is left in the learning model creation device Dd, the learning model creation method, and the learning model creation program, the redundant feature amount extraction unit is specified with priority given to maintaining the performance of the learning model. Will be done. When the input unit 1 is accepted to exclude the third feature amount extraction unit (blurring information extraction unit), the learning model creation device Dd, the learning model creation method, and the learning model creation program are used to extract the third feature amount. Since the redundant feature amount extraction unit is specified by excluding the unit, the third feature amount extraction unit can be left in the learning model after the redundant feature amount extraction unit is deleted. In this case, since the third feature quantity extraction unit is left in the learning model creation device Dd, the learning model creation method, and the learning model creation program, the redundant feature quantity extraction unit is specified with priority given to maintaining the performance of the learning model. Will be done.

FIG. 13 is a diagram for explaining the effect of the learning model creating device in the fourth embodiment. In one specific example, when creating a learning model for character recognition in which edge information is important without paying attention to the color difference of the input image, the third check box 723 is checked as shown in FIG. To. As a result, as shown in FIG. 13, the original learning model has a feature amount extraction unit A that reacts to yellow, a feature amount extraction unit B that reacts to the rising edge to the right, and a feature amount extraction unit that reacts to the edge rising to the left. C, feature amount extraction unit D that reacts to vertical edges, feature amount extraction unit E that reacts to horizontal edges, feature amount extraction unit F that reacts to light blue, feature amount extraction unit G that reacts to red, and curves. When the feature amount extraction unit M that reacts is included, the feature amount extraction unit that reacts to the edge information is removed from the redundant feature amount extraction unit specified, and the feature amount extraction unit A that reacts to yellow and the feature that reacts to light blue. The amount extraction unit F and the feature amount extraction unit G that reacts to red are deleted, and the feature amount extraction unit B that reacts to the rising edge, the feature amount extraction unit C that reacts to the left rising edge, and the feature amount extraction unit C that reacts to the left rising edge react to the vertical edge. A deleted learning model including the feature amount extraction unit D to be performed, the feature amount extraction unit E to react to the lateral edge, and the feature amount extraction unit M to react to the curve is created. As a result, a learning model according to the purpose is created by leaving the feature amount extraction unit that reacts to the edge.

Next, another embodiment will be described.

(Fifth Embodiment)
The learning model creating devices Da to Dc in the first to third embodiments specify the redundant feature quantity extraction unit with priority given to reduction of calculation cost, but the learning model creating device De in the fifth embodiment is learning. The redundant feature amount extraction unit is specified with priority given to maintaining the performance of the model.

As shown in FIG. 1, for example, the learning model creating device De in the fifth embodiment includes an input unit 1, an output unit 2, an IF unit 3, a control processing unit 4e, and a storage unit 5. Be prepared. In the input unit 1, the output unit 2, the IF unit 3, and the storage unit 5 in the learning model creation device De in the fifth embodiment, the input unit 1 further accepts priority items for prioritizing the maintenance of the performance of the learning model. Is the same as the input unit 1, the output unit 2, the IF unit 3 and the storage unit 5 in the learning model creating device Da in the first embodiment, and the description thereof will be omitted.

In the present embodiment, the input unit 1 further accepts the input of the second priority item that prioritizes the maintenance of the performance of the learning model, as described above. In the present embodiment, the second priority item is a first feature amount extraction unit (color information extraction unit) that extracts color information as the feature amount, and a second feature amount extraction unit that extracts edge information as the feature amount (the second feature amount extraction unit). A feature amount selected from the edge information extraction unit) and the third feature amount extraction unit (blur information extraction unit) that extracts blur information as the feature amount, and preferentially specified as the redundant feature amount extraction unit. It is an extraction part. The selected feature amount extraction unit may be one, may be plural, or may be all. The second priority item is another example of the specific condition. Therefore, the input unit 1 receives the input of the feature amount extraction unit that is preferentially specified as the redundant feature amount extraction unit from the first to third feature amount extraction units.

The control processing unit 4e is a circuit for creating a learning model for machine learning by controlling each unit 1 to 3 and 5 of the learning model creation device De according to the function of each unit. For example, the CPU and its surroundings. It is configured with a circuit. The control processing unit 4e functionally includes a control unit 41, a learning application unit 42, a specific unit 43e, and a model improvement unit 44 by executing the control processing program. The control unit 41, the learning application unit 42, and the model improvement unit 44 in the control processing unit 4e of the fifth embodiment are the control unit 41, the learning application unit 42, and the model improvement unit in the control processing unit 4a of the first embodiment, respectively. Since it is the same as 44, the description thereof will be omitted.

The specific unit 43e identifies a redundant feature amount extraction unit from a plurality of feature amount extraction units in the learning model during learning or when analyzing an input sample using the learning model after learning. It is something to do. In the present embodiment, the specific unit 43e specifies the redundant feature amount extraction unit with priority given to maintaining the performance of the learning model. More specifically, the specific unit 43e preferentially identifies the feature amount extraction unit corresponding to the feature amount extraction unit received by the input unit 1 as the redundant feature amount extraction unit, whereby the redundant feature is described. Identify the quantity extractor. That is, when the feature amount extraction unit in the learning model is the feature amount extraction unit corresponding to the feature amount extraction unit received by the input unit 1, the specific unit 43e has priority as the redundant feature amount extraction unit. Specific to. More specifically, for example, when the redundant feature amount extraction unit is specified in order from the feature amount extraction unit having the lowest contribution, the feature amount extraction unit corresponding to the feature amount extraction unit received by the input unit 1 contributes to this. It is removed from the plurality of feature amount extraction units arranged in order of degree and preferentially specified as the redundant feature amount extraction unit, and the specific unit 43d sets the feature amount extraction unit corresponding to the feature amount extraction unit received by the input unit 1. The redundant feature amount extraction unit is the number obtained by subtracting the number of feature amount extraction units preferentially specified as the redundant feature amount extraction unit from the number of deletion settings in order from the feature amount extraction unit with the lowest contribution. To identify. Further, for example, when the redundant feature amount extraction unit is specified in order from the feature amount extraction unit having the lower degree of independence, the feature amount extraction unit corresponding to the feature amount extraction unit received by the input unit 1 is in the order of the independence degree. It is removed from the plurality of feature amount extraction units arranged side by side and preferentially specified as the redundant feature amount extraction unit, and the specific unit 43d removes the feature amount extraction unit corresponding to the feature amount extraction unit received by the input unit 1. The redundant feature amount extraction unit is specified by the number obtained by subtracting the number of feature amount extraction units preferentially specified as the redundant feature amount extraction unit from the number of deletion settings in order from the feature amount extraction unit having the lowest degree of independence. do. Further, for example, a plurality of feature amount extraction units in the learning model are divided into a plurality of groups consisting of feature amount extraction units similar to each other, and residual feature amount extraction excluding the representative feature amount extraction unit is extracted from each of the plurality of groups. When the unit is specified as the redundant feature amount extraction unit, the feature amount extraction unit corresponding to the feature amount extraction unit received by the input unit 1 is removed from the plurality of feature amount extraction units before the grouping, and the redundancy unit is removed. The specific unit 43d is preferentially specified as a feature amount extraction unit, and the specific unit 43d divides the residual feature amount extraction unit into a plurality of groups consisting of feature amount extraction units similar to each other, and the representative from each of the plurality of groups. The remaining feature amount extraction unit excluding the feature amount extraction unit is specified as the redundant feature amount extraction unit. When the number of deletion settings is set, the number of feature quantity extraction units preferentially specified as the redundant feature quantity extraction unit is subtracted from the number of deletion settings, a new deletion setting quantity is obtained, and after grouping. The distance determination threshold is adjusted so that the total number of the groups is the total number of the feature amount extraction unit minus the new deletion setting number.

The learning model creating device De in such a fifth embodiment is shown in FIG. Each of the above-mentioned processes is carried out using 4. As a result, a learning model in which the redundant feature amount extraction unit is deleted is created.

On the other hand, in the analysis process, the learning model creation device De in the fifth embodiment performs the deletion target specific process S21e by the specific unit 43e instead of the deletion target specific process S21a by the specific unit 43a. Each of the above-mentioned treatments is carried out using. As a result, the analysis of the target sample is performed with a new learning model in which the redundant feature amount extraction unit is removed from the initial learning model.

In the deletion target specific processing S21e, the learning model creation device Dd receives the second priority item in the input unit 1 by the specific unit 43d, and receives the second priority item in the input unit 1 to the feature amount extraction unit of the second priority item. The redundant feature amount extraction unit is specified by preferentially specifying the corresponding feature amount extraction unit as the redundant feature amount extraction unit.

The learning model creation device De in the fifth embodiment, and the learning model creation method and the learning model creation program implemented therein are the learning model creation device Da, the learning model creation method, and the learning model in the first embodiment. It has the same effect as the created program.

Then, the learning model creation device De, the learning model creation method, and the learning model creation program use the learning model by designating the feature amount extraction unit that becomes noise according to the usage of the learning model as the second priority item. The performance according to the application can be maintained, and the performance of the learning model can be maintained by leaving an important feature amount extraction unit according to the purpose of the learning model according to the user's intention. When the input unit 1 is accepted to preferentially specify the first feature amount extraction unit (color information extraction unit), the learning model creation device De, the learning model creation method, and the learning model creation program are the first. Since the redundant feature amount extraction unit is specified by giving priority to the feature amount extraction unit, the learning model after deleting the redundant feature amount extraction unit can not include the first feature amount extraction unit. In particular, when the first feature quantity extraction unit becomes noise from the viewpoint of the usage of the learning model, the learning model creation device De, the learning model creation method, and the learning model creation program delete the redundant feature quantity extraction unit. The first feature amount extraction unit, which becomes noise, can be deleted (excluded) from the later learning model. In this case, in the learning model creation device De, the learning model creation method, and the learning model creation program, the first feature quantity extraction unit is preferentially deleted, so that the redundant features are prioritized to maintain the performance of the learning model. The quantity extractor is specified. When the input unit 1 is accepted to preferentially specify the second feature amount extraction unit (edge information extraction unit), the learning model creation device De, the learning model creation method, and the learning model creation program are second. Since the redundant feature amount extraction unit is specified by giving priority to the feature amount extraction unit, the learning model after deleting the redundant feature amount extraction unit can not include the second feature amount extraction unit. In particular, when the second feature quantity extraction unit becomes noise from the viewpoint of the usage of the learning model, the learning model creation device De, the learning model creation method, and the learning model creation program delete the redundant feature quantity extraction unit. The second feature amount extraction unit, which becomes noise, can be deleted from the later learning model. In this case, in the learning model creation device De, the learning model creation method, and the learning model creation program, the second feature quantity extraction unit is preferentially deleted, so that the redundant features are prioritized to maintain the performance of the learning model. The quantity extractor is specified. When the input unit 1 is accepted to preferentially specify the third feature amount extraction unit (blurring information extraction unit), the learning model creation device De, the learning model creation method, and the learning model creation program are the third. Since the redundant feature amount extraction unit is specified by giving priority to the feature amount extraction unit, the learning model after deleting the redundant feature amount extraction unit can not include the third feature amount extraction unit. In particular, when the third feature quantity extraction unit becomes noise from the viewpoint of the usage of the learning model, the learning model creation device De, the learning model creation method, and the learning model creation program delete the redundant feature quantity extraction unit. From the later learning model, the third feature amount extraction unit that becomes noise can be deleted. In this case, in the learning model creation device De, the learning model creation method, and the learning model creation program, the third feature quantity extraction unit is preferentially deleted, so that the redundant features are prioritized to maintain the performance of the learning model. The quantity extractor is specified.

In the above-described embodiment, the number of deletion settings may be set as follows. The total amount of the feature amount extraction unit to be deleted in the original learning model is b [Byte], and the number of calculation steps required to process the feature amount extraction unit of 1 [Byte] is S [Step / Byte]. When the processing time required to process the 1 [step] is t [s / step] and the required processing time in actual application is Rt [s], the accuracy required for the learning model can be obtained. In consideration, the number of deletion settings is set so that the deletion rate α is α × b × S × t ≦ Rt.

In order to express the present invention, the present invention has been appropriately and sufficiently described through the embodiments with reference to the drawings described above, but those skilled in the art can easily change and / or improve the above-described embodiments. It should be recognized that it is possible. Therefore, unless the modified or improved form implemented by a person skilled in the art is at a level that deviates from the scope of rights of the claims stated in the claims, the modified form or the improved form is the scope of rights of the claims. It is interpreted to be included in.

Da, Db, Dc, Dd, De Learning model creation device 1 Input unit 2 Output unit 3 Interface unit (IF unit)
4a, 4b, 4c, 4d, 4e Control processing unit 5 Storage unit 41 Control unit 42 Learning application unit 43a, 43b, 43c, 43d, 43e Specific unit 44 Model improvement unit 51 Teacher set storage unit 52 Target sample storage unit 511 Teacher data Storage unit 512 Correct label storage unit

Claims (3)

A learning model creation device that generates learning models for machine learning.
The learning model includes a plurality of feature quantity extraction units that acquire a function of extracting a predetermined feature quantity by learning.
A specific unit that identifies a redundant feature quantity extraction unit from the plurality of feature quantity extraction units during training or when analyzing an input sample using a learning model after training.
A model improvement unit that creates a new learning model by deleting the feature amount extraction unit specified by the specific unit from the learning model, and a model improvement unit.
A new learning model created by the model improvement unit, provided with a learning application unit for continuing the learning or analyzing the input sample.
The specific unit identifies the redundant feature amount extraction unit with priority given to reduction of calculation cost.
The specific unit identifies the redundant feature amount extraction unit in order from the feature amount extraction unit having a low degree of independence indicating the degree of independence between the feature amount extraction units.
Learning model creation device. A learning model creation device that generates learning models for machine learning.
The learning model includes a plurality of feature quantity extraction units that acquire a function of extracting a predetermined feature quantity by learning.
A specific unit that identifies a redundant feature quantity extraction unit from the plurality of feature quantity extraction units during training or when analyzing an input sample using a learning model after training.
A model improvement unit that creates a new learning model by deleting the feature amount extraction unit specified by the specific unit from the learning model, and a model improvement unit.
A new learning model created by the model improvement unit, provided with a learning application unit for continuing the learning or analyzing the input sample.
The specific unit identifies the redundant feature amount extraction unit with priority given to maintaining the performance of the learning model.
A first feature amount extraction unit that extracts color information as the feature amount, a second feature amount extraction unit that extracts edge information as the feature amount, and a third feature amount extraction unit that extracts blur information as the feature amount. Further, a first input unit for receiving an input of a feature amount extraction unit excluded from the specified redundant feature amount extraction unit is further provided.
The specific unit excludes the feature amount extraction unit corresponding to the feature amount extraction unit received by the first input unit from the redundant feature amount extraction unit to be specified, thereby removing the redundant feature amount extraction unit. Identify,
Learning model creation device. A learning model creation device that generates learning models for machine learning.
The learning model includes a plurality of feature quantity extraction units that acquire a function of extracting a predetermined feature quantity by learning.
A specific unit that identifies a redundant feature quantity extraction unit from the plurality of feature quantity extraction units during training or when analyzing an input sample using a learning model after training.
A model improvement unit that creates a new learning model by deleting the feature amount extraction unit specified by the specific unit from the learning model, and a model improvement unit.
A new learning model created by the model improvement unit, provided with a learning application unit for continuing the learning or analyzing the input sample.
The specific unit identifies the redundant feature amount extraction unit with priority given to maintaining the performance of the learning model.
A first feature amount extraction unit that extracts color information as the feature amount, a second feature amount extraction unit that extracts edge information as the feature amount, and a third feature amount extraction unit that extracts blur information as the feature amount. Further, a second input unit that accepts the input of the feature amount extraction unit that is preferentially specified as the redundant feature amount extraction unit is further provided.
The specific unit preferentially identifies the feature amount extraction unit corresponding to the feature amount extraction unit received by the second input unit as the redundant feature amount extraction unit, thereby providing the redundant feature amount extraction unit. Identify,
Learning model creation device.

Images