JP2024039138A - Model selection device, model selection method, and program - Google Patents

Abstract

[Problem] A model selection device that can improve the accuracy of a trained model that has undergone transfer learning by selecting a trained model that is optimal for adaptation to transfer learning from among multiple trained models with different domains. , a model selection method, and a program. [Solution] A domain learning unit that generates, for each different domain having a common class, a trained model trained to identify the identification target using input data indicating an identification target as input; an optimal model selection unit that selects a trained model that is optimal for adaptation to transfer learning from among the trained models that have been trained, based on an identification result in which each trained model has identified the identification target. Device. [Selection diagram] Figure 1

Description

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

In recent years, machine learning has been used to create classifiers that identify input data. When creating different classifiers for a large amount of input data and a small amount of input data, in normal machine learning, each classifier is created by machine learning using a large amount of input data and machine learning using a small amount of input data. do. In this case, a classifier created by machine learning using a small amount of input data tends to have low classification accuracy. For this reason, in recent years, transfer learning, which is a technique for adapting a classifier (trained model) trained using a large amount of input data to a small amount of input data, has been increasingly used.

In learning these input data, for example, a trained CNN (Convolution Neural Networks) can be obtained by inputting the input data into a model called a CRNN (Convolution Recurrent Neural Networks) and having the model learn the identification target.

Normally, when performing transfer learning, a trained model is created using a large amount of input data, and the trained model is reused. This allows transfer learning to perform learning using a small amount of labeled input data as input, and is expected to reduce learning time and improve accuracy in transfer learning.

However, when there are multiple trained models that can be reused in transfer learning, it is difficult to know in advance which trained model is most suitable for transfer learning (for example, which has the highest accuracy).

Furthermore, when the type of input data, the possible values of the input value, the type of output data, the possible values of the output, the probability distribution, etc. are different (that is, the domain is different), the same model can be applied to the trained model. Reusable technology has been established. However, there has been no method for determining the domain and selecting an appropriate trained model for different types of recognition targets (for example, handwritten characters written in different years).

Patent Document 1 listed below discloses a technique for selecting a specific classifier group based on a vector dispersion relationship in order to evaluate a classifier group (weak classifier group).
Further, Patent Document 2 listed below discloses a technology for translating OCR text read from a source language by OCR (Optical Character Recognition) into a target language. With this technology, translated OCR text can be obtained by determining the complexity of translation and performing machine language translation of the OCR text from the source language to the target language based on the complexity.

JP2019-191711A US Patent No. 9514377

By the way, in a classifier, the domain in which the classification accuracy is high may differ depending on the input data. In this case, when applying transfer learning, it is desirable to determine in advance in which domain the classification accuracy will be high, and to use the classifier with the highest classification accuracy for transfer learning. However, with the techniques of Patent Document 1 and Patent Document 2, it is difficult to consider the identification accuracy for each domain in advance.

In view of the above-mentioned problems, the purpose of the present invention is to improve the accuracy of a trained model that has undergone transfer learning by selecting a trained model that is optimal for adaptation to transfer learning from trained models of different domains. The object of the present invention is to provide a model selection device, a model selection method, and a program that can perform the following tasks.

In order to solve the above-mentioned problem, a model selection device according to one aspect of the present invention learns to identify the identification target using input data indicating the identification target for each different domain having a common class. A domain learning unit that generates a trained model, and a trained model that is optimally adapted to transfer learning based on the identification result of identifying the classification target from among the trained models generated for each of the different domains. and an optimal model selection unit that selects a learned model.

In the model selection method according to one aspect of the present invention, the domain learning unit selects, for each different domain having a common class, a trained model that has been trained to identify the identification target using input data indicating the identification target as input. The domain learning process to be generated and the optimal model selection unit adapt to transfer learning based on the identification results of each trained model identifying the identification target from among the trained models generated for each of the different domains. an optimal model selection step of selecting a trained model optimal for the method.

A program according to an aspect of the present invention performs domain learning that causes a computer to generate a trained model that has been trained to identify the identification target by inputting input data indicating an identification target for each different domain having a common class. and an optimal method for selecting a trained model that is optimal for adaptation to transfer learning from among the trained models generated for each of the different domains, based on the identification result in which each trained model identifies the identification target. Function as a model selection means.

According to the present invention, by selecting a trained model that is optimal for adaptation to transfer learning from trained models of different domains, it is possible to improve the accuracy of a trained model that has been subjected to transfer learning.

1 is a block diagram showing an example of a functional configuration of a model selection device according to a first embodiment. FIG. FIG. 2 is a flowchart illustrating an example of a process for generating a trained model that is a transfer learning candidate according to the first embodiment; FIG. 2 is a flowchart illustrating an example of the flow of a learned model selection process according to the first embodiment. FIG. 2 is a block diagram illustrating an example of a functional configuration of a model selection device according to a second embodiment. 7 is a flowchart illustrating an example of the flow of a learned model selection process according to the second embodiment.

Embodiments of the present invention will be described in detail below with reference to the drawings.

<1. First embodiment>
A first embodiment will be described with reference to FIGS. 1 to 3.
In the first embodiment, each trained model is selected from among the trained models that have been trained to identify the classification target for each classifier that has a different target domain for the classifier that identifies the classification target included in the input data. We will explain a model selection device that selects a trained model that is optimal for adaptation to transfer learning based on the domain score of . It is assumed that each domain has a common class.

In the following, an example will be described in which the identification target is a handwritten character and the input data is image data representing the handwritten character.
The domain in this case is, for example, a type of classical book or ancient document that includes text information. Each domain is classified according to this type, and it is assumed that data of each type is not mixed in each domain.
The class is, for example, a character type. Character types are classifications such as kanji, hiragana, and katakana. As an example, domains that include characters expressed in Kanji are classified into the same class, and can be said to have a common class. The same applies to hiragana and katakana. Further, the class may be a font of characters. The font of the characters is classified into, for example, regular script, cursive script, and cursive script. As an example, domains that include characters shown in block font are classified into the same class, and can be said to have a common class. The same applies to Gyosho and Cursive.

Further, a trained model when the recognition target is a handwritten character is generated for each domain by learning using a neural network in a handwriting OCR engine. Note that OCR is an abbreviation for Optical Character Recognition, and refers to a technology that recognizes text in an image and converts it into text data.
After generation, a domain score is calculated for each trained model based on an output obtained by inputting image data of a handwritten region including text data. After calculation, the trained model that is most suitable for transfer learning is selected from among the trained models based on the calculated domain score.

Furthermore, it is assumed that the handwritten characters to be identified by each of the trained models generated for each domain are written in different years for each trained model. Further, it is assumed that the learned model is a model learned using image data showing handwritten characters written with a writing instrument.
Further, when image data is provided as a line image, the model selection device handles as input data an image in line units (for example, one line unit) divided at each domain break. On the other hand, if the input data is not provided as a line image, the model selection device may treat an image of each character separated by character as the input data.

<1-1. Functional configuration of model selection device>
With reference to FIG. 1, the functional configuration of the model selection device according to the first embodiment will be described. FIG. 1 is a block diagram showing an example of a functional configuration of a model selection device according to a first embodiment. As shown in FIG. 1, the model selection device 10 includes a storage section 110, a model analysis section 120, a model evaluation section 130, and an optimal model selection section 140.

(1) Storage unit 110
The storage unit 110 has a function of storing various information. The storage unit 110 includes a storage medium such as an HDD (Hard Disk Drive), an SSD (Solid State Drive), a flash memory, an EEPROM (Electrically Erasable Programmable Read Only Memory), or a RAM ( Random Access read/write Memory), ROM (Read Only Memory), or any combination of these storage media.
The storage unit 110 may include a database (DB) for storing various information. For example, as shown in FIG. 1, the storage unit 110 includes an input image DB 111, a model management DB 112, and a calculation result DB 113.

(1-1) Input image DB111
The input image DB 111 is a database that stores input data used as input when generating a trained model and input data used as input when obtaining an output from the generated trained model.

(1-2) Model management DB112
The model management DB 112 is a trained model generated by a domain learning unit 121 described later, and stores trained models that are candidates to which transfer learning is applied.

(1-3) Calculation result DB113
The calculation result DB 113 stores an index for selecting a trained model that is optimal for applying transfer learning. The index is, for example, a domain score calculated by the model evaluation unit 130 described later.

(2) Model analysis section 120
The model analysis unit 120 has a function of analyzing image data including text information extracted from the input image DB 111 and extracting trained models that are candidates for application to transfer learning.
As shown in FIG. 1, the model analysis section 120 includes a domain learning section 121.

(2-1) Domain learning section 121
The domain learning unit 121 has a function of generating a trained model for each domain. For example, the domain learning unit 121 acquires image data including text information from the input image DB 111, and uses each image data as input to discriminate (identify) and learn character types, thereby generating a learned model for each domain. and output it. The generated trained model is stored in the model management DB 112.

Note that the domain learning unit 121 may have correspondence map data indicating the correspondence between different classes. In this case, the domain learning unit 121 performs a process of integrating classes between trained models for the selected class based on the corresponding map data. As an example, when the correspondence map data shows the correspondence between kanji and hiragana, for example, the identification result of a kanji whose reading is "a" and the identification result of the hiragana "a" can be considered to be the same. Thereby, the domain learning unit 121 can treat domains having different classes as domains having a common class.

(3) Model evaluation section 130
The model evaluation unit 130 has a function of evaluating a learned model. For example, the model evaluation unit 130 calculates the variance for each trained model in a different domain, and evaluates each trained model based on the calculation results.
As shown in FIG. 1, the model evaluation unit 130 includes a probability distribution calculation unit 131, an information matrix calculation unit 132, and a domain score calculation unit 133.

(3-1) Probability distribution calculation unit 131
The probability distribution calculation unit 131 has a function of calculating a probability distribution. For example, the probability distribution calculation unit 131 receives the trained model obtained by the domain learning unit 121 as input and calculates the probability distribution of image data (input data) that is a target of transfer learning. The probability distribution calculation unit 131 calculates probability distributions for all trained models for each domain.

(3-2) Information matrix calculation unit 132
The information matrix calculation unit 132 has a function of calculating an information matrix. For example, the information matrix calculation unit 132 receives the probability distribution calculated by the probability distribution calculation unit 131 as an input, uses the probability distribution as a vector, and generates an information matrix by calculating a gradient.

(3-3) Domain score calculation unit 133
The domain score calculation unit 133 has a function of calculating a domain score. The domain score calculation unit 133 calculates the variance of the identification results for identifying handwritten characters as a domain score for each trained model generated by the domain learning unit 121, as an index for selecting an optimal trained model. For example, the domain score calculation unit 133 receives the information matrix calculated by the information matrix calculation unit 132 as input and calculates a domain score for each domain. The calculated domain score is stored in the calculation result DB 113.

(4) Optimal model selection unit 140
The optimal model selection unit 140 has a function of selecting a trained model that is optimal for adaptation to transfer learning. For example, the optimal model selection unit 140 selects a handwritten character from among the trained models generated by the domain learning unit 121 and stored in the model management DB 112 based on the identification result of each trained model identifying handwritten characters. Select the best trained model for adaptation. In the first embodiment, the optimal model selection unit 140 selects the optimal trained model from the model management DB 112 based on the domain score calculated by the domain score calculation unit 133 and stored in the calculation result DB 113.

<1-2. Processing flow>
The functional configuration of the model selection device 10 according to the first embodiment has been described above. Next, the flow of processing according to the first embodiment will be described with reference to FIGS. 2 and 3. Below, as an example, an example will be described in which an OCR system that handles image data including text information as input outputs a trained model and then uses the optimal trained model for new input data. Note that the image data including text information specifies the ID, width, height, x coordinate to specify the starting position of the width of the line image, ID, width, height, and starting position of the height of the line image to identify the line image. It is assumed that there is a character string in one image data that spans multiple lines, including the y-coordinate for the image, character string attributes, etc.

(1) Generation process of a trained model to be a transfer learning candidate First, an example of the flow of generation process for a trained model to be a transfer learning candidate will be described with reference to FIG. FIG. 2 is a flowchart illustrating an example of the process of generating a trained model that is a transfer learning candidate according to the first embodiment.

As shown in FIG. 2, the domain learning unit 121 first obtains image data including text data from the input image DB 111 (step S101).
Next, the domain learning unit 121 checks whether the acquired image data has been binarized (step S102). If it has not been binarized (step S102/NO), the process advances to step S103. On the other hand, if the data has been binarized (step S102/YES), the process advances to step S104.

When the process proceeds to step S103, the domain learning unit 121 binarizes the image data (step S103). For example, the domain learning unit 121 performs a discriminant separation method to find a threshold value that maximizes the degree of separation, and automatically binarizes the image. After the binarization, the process advances to step S104.

When the process proceeds to step S104, the domain learning unit 121 cuts out the image data into rectangular images for each image (step S104). For example, the domain learning unit 121 calculates the circumscribing rectangle of one line from the width, height, x coordinate for specifying the starting position of the width of the line image, and y coordinate for specifying the starting position of the height of the line image. A rectangular image is generated based on the specified circumscribed rectangle. Then, the domain learning unit 121 outputs a character string corresponding to each rectangular image.

Next, the domain learning unit 121 checks whether it is necessary to create label information in learning (step S105). If label information is not prepared, it is determined that label information needs to be created (step S105/YES), and the process proceeds to step S106. On the other hand, if label information is prepared, it is determined that creation of label information is not necessary (step S105/NO), and the process proceeds to step S107.

When the process proceeds to step S106, the domain learning unit 121 creates label information (step S106). In the first embodiment, character type semantic segmentation (SS) is created for learning. In the character type SS, each character type is labeled as a correct answer and used for the class. After creating the label information, the process advances to step S107.

When the process proceeds to step S107, the domain learning unit 121 estimates the character type using the rectangular image based on the label information for each domain (step S107).
Next, the domain learning unit 121 performs learning using a neural network for estimating handwritten portions using label information for each domain (step S108). These network structures may take the form of FCN (Fully Convolutional Networks), for example.
Then, the domain learning unit 121 outputs the trained model generated for each domain, and stores and saves it in the model management DB 112 (step S109).

(2) Flow of the trained model selection process Next, an example of the flow of the trained model selection process will be described with reference to FIG. FIG. 3 is a flowchart illustrating an example of the flow of a trained model selection process according to the first embodiment.

As shown in FIG. 3, the model evaluation unit 130 first obtains rectangular image data (step S201). For example, the model evaluation unit 130 acquires binarized rectangular image data including text data in the same manner as steps S101 to S104 in FIG. 2 .

Next, the model evaluation unit 130 checks whether there are multiple trained models for each domain stored in the model management DB 112 (step S202). If there are multiple trained models for one domain (step S202/YES), the process advances to step S203. On the other hand, if there are not a plurality of trained models for one domain (step S202/NO), the process advances to step S204.

When the process proceeds to step S203, the model evaluation unit 130 selects the most suitable trained model within the domain from among the plurality of trained models (step S203). After selection, the process advances to step S204. Note that if a domain in which no trained model is stored is detected in step S202, the trained model is excluded from domain candidates when selecting the optimal trained model.

When the process proceeds to step S204, the probability distribution calculation unit 131 calculates a probability distribution from the output of the learned model stored in the model management DB 112 (step S204). For example, the probability distribution calculation unit 131 calculates a dummy label distribution of the target data set using the learned model as input.

Next, the information matrix calculation unit 132 calculates an information matrix (step S205). For example, the information matrix calculation unit 132 calculates the gradient for each parameter of the probability distribution calculated as a dummy label distribution by the probability distribution calculation unit 131, and calculates the slope of the probability distribution from the input data and the probability distribution by multiplying the gradient with the transposed matrix. Calculate the information matrix.

Next, the information matrix calculation unit 132 checks whether there is any overlap in the classes for each domain (step S206). If there is no overlap (step S206/NO), the process advances to step S207. On the other hand, if there is overlap (step S206/YES), the process advances to step S208.

When the process proceeds to step S207, the information matrix calculation unit 132 performs filtering of the information matrix (step S207). In this filtering, all filter parameters are averaged, assuming that the correlation between different filters is not important from the information matrix. This allows simplification by performing approximation from a complete information matrix. After filtering, the process advances to step S208.

When the process proceeds to step S208, the domain score calculation unit 133 outputs a fixed-length vector from the information matrix for each domain (step S208). In the first embodiment, this fixed-length vector corresponds to the domain score. Since the size of the information matrix differs for each trained model, the domain score calculation unit 133 extracts only the diagonal components, averages the values of the diagonal components with the same filter, and outputs a fixed-length vector.

Next, the domain score calculation unit 133 checks the symmetry of each learned model (step S209). If there is no symmetry (asymmetric) (step S209/NO), the process advances to step S210. On the other hand, if there is symmetry (step S209/YES), the process advances to step S211.

When the process proceeds to step S210, the domain score calculation unit 133 calculates a domain score for each domain using an asymmetric degree of similarity (step S210), and then proceeds to step S212.
When the process proceeds to step S211, the domain score calculation unit 133 vectorizes the trained model using cosine similarity and calculates a domain score (step S211). After the calculation, the process advances to step S212.
Note that the calculated domain score is stored and saved in the calculation result DB 113.

When the process proceeds to step S212, the optimal model selection unit 140 selects the optimal trained model based on the domain score calculation result (step S212). For example, in calculating the domain score for each domain, the optimal model selection unit 140 selects the trained model showing the smallest score as the optimal trained model.

As explained above, the model selection device 10 according to the first embodiment uses a trained model that has been trained to identify an identification target by inputting input data indicating an identification target for each different domain having a common class. The domain learning unit 121 generates a trained model that is optimal for adaptation to transfer learning based on the identification results of each trained model identifying the classification target from among the trained models generated for each different domain. and an optimal model selection unit 140 that selects the optimal model.

With this configuration, the model selection device 10 according to the first embodiment can select a trained model that is suitable for transfer learning, taking into consideration in advance the difference in the classification accuracy of the classifier due to the difference in domains. .
Therefore, the model selection device 10 according to the first embodiment improves the accuracy of the trained model that has undergone transfer learning by selecting the trained model that is optimal for adaptation to transfer learning from among the trained models of different domains. This makes it possible to improve

The model selection device 10 according to the first embodiment also performs domain score calculation that calculates the variance of identification results for each trained model generated for each different domain, as an index for selecting an optimal trained model. The optimal model selection unit 140 selects the optimal trained model based on the calculated domain score.
With this configuration, the model selection device 10 according to the first embodiment selects the optimal model for inference from the input data by using the true value of the input data and the probability distribution of the learned model of the input data. It is possible to select with higher accuracy. Furthermore, the time required to select the optimal model can also be reduced.

<2. Second embodiment>
The first embodiment has been described above. Next, a second embodiment will be described with reference to FIGS. 4 and 5.
In the first embodiment described above, an example was described in which an optimal trained model is selected from trained models that are candidates for transfer learning based on the domain score of each trained model, but the present invention is limited to this example. Not done. In the second embodiment, an example will be described in which an optimal trained model is selected based on ensemble learning for trained models that are candidates for transfer learning.
Note that, hereinafter, explanations that overlap with those in the first embodiment will be omitted as appropriate.

<2-1. Functional configuration of model selection device>
With reference to FIG. 4, the functional configuration of a model selection device 10a according to the second embodiment will be described. FIG. 4 is a block diagram showing an example of the functional configuration of a model selection device 10a according to the second embodiment. As shown in FIG. 4, the model selection device 10a includes a storage section 110a, a model analysis section 120a, and an optimal model selection section 140a.

(1) Storage unit 110a
The storage unit 110a has a function of storing various information using a storage medium similar to that of the storage unit 110 according to the first embodiment. As shown in FIG. 4, the storage unit 110a includes an input image DB111, a model management DB112a, and a calculation result DB113.

(1-1) Input image DB111
The functions of the input image DB 111 according to the second embodiment are the same as the functions of the input image DB 111 according to the first embodiment, so the description thereof will be omitted.

(1-2) Model management DB112a
The model management DB 112a is a trained model generated by a transfer learning unit 122 described later, and stores trained models to which ensemble learning is applied by an ensemble learning unit 123 described later.

(1-3) Calculation result DB113
The functions of the calculation result DB 113 according to the second embodiment are the same as those of the calculation result DB 113 according to the first embodiment, so the description thereof will be omitted.

(2) Model analysis section 120a
As shown in FIG. 4, the model analysis unit 120a further includes a transfer learning unit 122 and an ensemble learning unit 123 in addition to a domain learning unit 121 similar to the model analysis unit 120 according to the first embodiment.

(2-1) Domain learning section 121
The functions of the domain learning unit 121 according to the second embodiment are the same as those of the domain learning unit 121 according to the first embodiment, so a description thereof will be omitted.

(2-2) Transfer learning section 122
The transfer learning unit 122 has a function of performing transfer learning on a trained model. For example, the transfer learning unit 122 performs transfer learning on each of the learned models generated by the domain learning unit 121 using target image data for the transfer destination domain.

(2-3) Ensemble learning section 123
The ensemble learning unit 123 has a function of performing ensemble learning based on trained models generated for different domains. For example, the ensemble learning unit 123 receives as input the trained model that has been transferred learned by the transfer learning unit 122, and performs ensemble learning on the trained model.

The ensemble learning unit 123 uses an arbitrary activation function to calculate weights as a distributed representation in the trained models in learning the trained models to be ensembled when performing ensemble learning. The ensemble learning unit 123 uses, for example, a softmax function as an activation function, and calculates the weight of the ensemble based on the softmax value. Note that the activation function used by the ensemble learning unit 123 is not limited to a softmax function as long as it is a function that converts and outputs an output value of 1.0, and may be, for example, a sigmoid function or a ReLU ( Other activation functions may be used, such as a Rectified Linear Unit) function.

(3) Optimal model selection unit 140a
The optimal model selection unit 140a, like the optimal model selection unit 140a according to the first embodiment, selects each trained model from among the trained models generated by the domain learning unit 121 and stored in the model management DB 112. Based on the recognition results of handwritten characters, the trained model that is most suitable for transfer learning is selected. In the second embodiment, the optimal model selection unit 140a selects the optimal trained model from the model management DB 112 based on the result of ensemble learning by the ensemble learning unit 123.

<2-2. Processing flow>
The functional configuration of the model selection device 10a according to the second embodiment has been described above. Next, the flow of processing according to the second embodiment will be described with reference to FIG. 5. In the following, similar to the first embodiment, for an OCR system that handles image data including text information as input, after outputting a trained model, the optimal trained model is used for new input data. Let's discuss an example. Note that the image data including text information specifies the ID, width, height, x coordinate to specify the starting position of the width of the line image, ID, width, height, and starting position of the height of the line image to identify the line image. It is assumed that there is a character string in one image data that spans multiple lines, including the y-coordinate for the image, character string attributes, etc.

(1) Generation process of a trained model that becomes a transfer learning candidate The process of generating a trained model that becomes a transfer learning candidate according to the second embodiment is the process described with reference to FIG. 2 in the first embodiment. Since it is the same as , its explanation will be omitted.

(2) Flow of the trained model selection process Next, an example of the flow of the trained model selection process will be described with reference to FIG. FIG. 5 is a flowchart illustrating an example of the flow of a trained model selection process according to the second embodiment.

The processing from step S301 to step S303 shown in FIG. 5 is the same as the processing from step S201 to step S303 described with reference to FIG. 3 in the first embodiment, so the description thereof will be omitted.

When the process proceeds to step S304, the transfer learning unit 122 performs transfer learning on all trained models generated by the domain learning unit 121 using image data (input data) (step S304). The learned model (transfer model) that has undergone transfer learning is stored and saved in the model management DB 112a.

Next, the ensemble learning unit 123 checks whether an algorithm is used for weighting (step S305). When using the algorithm (step S305/YES), the process advances to step S306. On the other hand, if the algorithm is not used (step S305/NO), the process advances to step S307.

When the process proceeds to step S306, the ensemble learning unit 123 calculates the probability of label estimation (step S306). For example, the ensemble learning unit 123 calculates the probability of label estimation using an arbitrary activation function. Specifically, the ensemble learning unit 123 calculates the probability of label estimation as an overall output by averaging the calculated probability of label estimation for each trained model and standardizing it with an arbitrary activation function. . After the calculation, the process advances to step S311.

When the process proceeds to step S307, the ensemble learning unit 123 acquires a label for each trained model (step S307).
Next, the ensemble learning unit 123 takes a majority vote based on the number of analogy results for the estimation results of the labels according to the number of acquired labels (step S308). The ensemble learning unit 123 determines a label with a large majority result as an analogous label. At this time, the ensemble learning unit 123 allocates IDs to the number of analogy results in descending order, and stores the numbers in the calculation result DB 113.

Next, the ensemble learning unit 123 checks whether the numbers of analogies in the majority voting result match (step S309). If they match (step S309/YES), the process advances to step S310. On the other hand, if they do not match (step S309/NO), the process advances to step S311.
When the process proceeds to step S310, the ensemble learning unit 123 selects the label of the analogy result with the smaller assigned ID (step S310). After selection, the process advances to step S311.

When the process proceeds to step S311, the ensemble learning unit 123 performs ensemble learning and outputs the result (step S311). When an algorithm is not used for weighting, the ensemble learning unit 123 uses a majority vote of the labels as the aggregation function and outputs ensemble accuracy. On the other hand, when using an algorithm for weighting, the ensemble learning unit 123 uses the aggregation function as an average of probabilities and outputs ensemble accuracy.
Then, the optimal model selection unit 140a selects the trained model in the ensemble as the optimal trained model (step S312).

As described above, the model selection device 10a according to the second embodiment uses a trained model that is trained to identify an identification target by inputting input data indicating an identification target for each different domain having a common class. The domain learning unit 121 generates a trained model that is optimal for adaptation to transfer learning based on the identification results of each trained model identifying the classification target from among the trained models generated for each different domain. and an optimal model selection unit 140a that selects the optimal model.

With this configuration, the model selection device 10a according to the second embodiment can select a trained model that is suitable for transfer learning, taking into consideration in advance the difference in the classification accuracy of the classifier due to the difference in domains. .
Therefore, the model selection device 10a according to the second embodiment improves the accuracy of the trained model that has undergone transfer learning by selecting the trained model that is optimal for adaptation to transfer learning from among the trained models of different domains. This makes it possible to improve

The model selection device 10a according to the second embodiment further includes an ensemble learning unit 123 that performs ensemble learning based on trained models generated for each different domain, and the optimal model selection unit 140 performs ensemble learning. Select the best trained model based on the results.
With this configuration, the model selection device 10a according to the second embodiment evaluates the true value of the item sought by the input data and the error of the target variable obtained by inputting the input data to the trained model and inferring it. By doing so, the optimal model for inference from input data can be selected with higher accuracy. Furthermore, the time required to select the optimal model can also be reduced.

The embodiments of the present invention have been described above. Note that some or all of the functions of the model selection devices 10 and 10a in the embodiments described above may be realized by a computer. In that case, a program for realizing this function may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read into a computer system and executed. Note that the "computer system" herein includes hardware such as an OS and peripheral devices. Furthermore, the term "computer-readable recording medium" refers to portable media such as flexible disks, magneto-optical disks, ROMs, and CD-ROMs, and storage devices such as hard disks built into computer systems. Furthermore, a "computer-readable recording medium" refers to a storage medium that dynamically stores a program for a short period of time, such as a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. It may also include a device that retains a program for a certain period of time, such as a volatile memory inside a computer system that is a server or client in that case. Further, the above-mentioned program may be one for realizing a part of the above-mentioned functions, or may be one that can realize the above-mentioned functions in combination with a program already recorded in the computer system. It may be realized using a programmable logic device such as an FPGA (Field Programmable Gate Array).

Although the embodiments of this invention have been described above in detail with reference to the drawings, the specific configuration is not limited to that described above, and various design changes may be made without departing from the gist of this invention. It is possible to do so.

DESCRIPTION OF SYMBOLS 10, 10a... Model selection device, 110, 110a... Storage unit, 111... Input image DB, 112, 112a... Model management, 113... Calculation result DB, 120, 120a... Model analysis unit, 121... Domain learning unit, 122... Transfer learning section, 123... Ensemble learning section, 130... Model evaluation section, 131... Probability distribution calculation section, 132... Information matrix calculation section, 133... Domain score calculation section, 140, 140a... Optimal model selection section

Claims (11)

a domain learning unit that generates a trained model trained to identify the identification target by inputting input data indicating the identification target for each different domain having a common class;
an optimal model selection unit that selects a trained model that is optimal for adaptation to transfer learning from among the trained models generated for each of the different domains, based on the identification result in which each trained model identifies the identification target; and,
A model selection device comprising: a domain score calculation unit that calculates the variance of the identification results for each trained model generated for each different domain as an index for selecting the optimal trained model;
Furthermore,
The optimal model selection unit selects the optimal trained model based on the calculated variance;
The model selection device according to claim 1. an ensemble learning unit that performs ensemble learning based on the trained models generated for each of the different domains;
Furthermore,
The optimal model selection unit selects the optimal trained model based on the result of the ensemble learning.
The model selection device according to claim 1. a transfer learning unit that performs transfer learning on a transfer destination domain for each of the trained models generated for each of the different domains;
Furthermore,
The ensemble learning unit performs the ensemble learning on the trained model that has been subjected to transfer learning.
The model selection device according to claim 3. The ensemble learning unit calculates a weight as a distributed representation in the trained model using an arbitrary activation function in learning the trained model to be ensembled when performing the ensemble learning.
The model selection device according to claim 3. The ensemble learning unit calculates the weight of the ensemble based on the softmax value.
The model selection device according to claim 5. The input data is image data indicating characters to be identified,
The trained models generated for each domain have different eras in which the characters to be identified were written, and are models trained using image data showing characters written with a writing instrument.
The model selection device according to claim 1. The domain learning unit has correspondence map data indicating a correspondence relationship between different classes, and performs a process of integrating classes between trained models on the selected class based on the correspondence map data.
The model selection device according to claim 1. When the input data is given as a row image, a row-by-row image divided at each domain break is treated as input data, and when the input data is not given as a row image, the input data is treated as input data. handles character-based images as input data,
The model selection device according to claim 1. a domain learning process in which the domain learning unit receives input data indicating an identification target and generates a trained model trained to identify the identification target for each different domain having a common class;
The optimal model selection unit selects a trained model that is optimal for adaptation to transfer learning from among the trained models generated for each of the different domains, based on the identification result of each trained model identifying the identification target. an optimal model selection process;
Model selection methods, including: computer,
domain learning means for generating a trained model trained to identify the identification target by inputting input data indicating the identification target for each different domain having a common class;
Optimal model selection means for selecting a trained model optimal for adaptation to transfer learning from among the trained models generated for each of the different domains, based on an identification result in which each trained model identifies the identification target; and,
A program to function as

Images