WO2024034196A1

WO2024034196A1 - Trained model selection method, trained model selection device, and trained model selection program

Info

Publication number: WO2024034196A1
Application number: PCT/JP2023/016265
Authority: WO
Inventors: 将隆窪内; 拓磨西本
Original assignee: 堺化学工業株式会社
Priority date: 2022-08-09
Filing date: 2023-04-25
Publication date: 2024-02-15
Also published as: JP2024024504A

Abstract

A trained model selection method for selecting a trained model for analyzing an analysis target from among a plurality of trained models M1 to Mn, said trained model selection method selecting the trained model for the analysis target using model language vectors V1 to Vn, which are language vectors corresponding to the trained models M1 to Mn respectively, model feature quantities F1 to Fn, which are feature quantities of the trained models M1 to Mn respectively, or the degree of similarity between the analysis target and the training data that has been used for machine learning of each of the trained models M1 to Mn.

Description

Trained model selection method, trained model selection device, and trained model selection program

The present invention relates to a technique for selecting a trained model for analyzing an analysis target from a plurality of trained models.

In analyzing data such as images, trained models generated by deep learning etc. have been used to analyze data. When it is important to be able to deal with all types of tasks, it is preferable to generate a trained model with generalization performance, but in general, compared to a specialized trained model, a trained model with high generalization performance is suitable for each task. accuracy is poor. Therefore, multiple trained models are often generated that are specialized for the task.

Previously, when there were multiple trained models, it was up to humans to decide which trained model to use for the task, requiring trial and error. Therefore, it took time to select the optimal trained model.

Therefore, a technology has been proposed that automatically selects the optimal trained model for the task. For example, in Patent Document 1, when data (character data in image data, etc.) is input, similar test data with the highest degree of similarity to the input data is calculated from a similar database prepared in advance, and the similar test data is It is disclosed that the trained model is processed using a plurality of trained models and the trained model with the highest accuracy rate is selected. In addition, in Patent Document 2, test data (user's face image, etc.) with attribute information (user's age, gender, etc.) is processed by multiple trained models prepared in advance, and the trained model with the highest accuracy rate is used. It is disclosed that it is possible to select.

JP 2019-40417 Publication International Publication No. 2018/142766

However, in the conventional techniques described in Patent Documents 1 and 2, it is necessary to input test data to a plurality of prepared trained models and evaluate the accuracy rate (performance) of each trained model, so the calculation process There is a problem that the amount of training models increases and it takes time to select a trained model.

The present invention was made to solve the above problem, and an object of the present invention is to quickly select a trained model suitable for an analysis target from a plurality of trained models.

A trained model selection method according to the present invention is a trained model selection method for selecting a trained model for analyzing an analysis target from a plurality of trained models, the method comprising: a trained model corresponding to each of the plurality of trained models; A model language vector that is a language vector, a model feature that is a feature of each of the plurality of trained models, or a link between the training data used for machine learning of each of the plurality of trained models and the analysis target. A trained model corresponding to the analysis target is selected using the degree of similarity.

According to a preferred embodiment, the model language vector is generated from language labels assigned to each of the trained models.

According to a preferred embodiment, the model language vector is generated from a language label given to teacher data used for machine learning of each of the trained models.

According to a preferred embodiment, the learned model selection method includes a receiving step of accepting the analysis target or language data corresponding to the analysis target as search data, and converting the search data into a language vector into a search language vector. a language vector conversion step, a language vector comparison step of comparing the search language vector with each of the model language vectors, and learning at least one from the plurality of trained models based on the comparison results of the language vector comparison step. a selection step of selecting a completed model.

According to a preferred embodiment, in the selection step, a trained model corresponding to a model language vector having the highest degree of similarity to the search language vector is selected.

According to a preferred embodiment, the learned model selection method includes a receiving step of accepting the analysis target or language data corresponding to the analysis target as search data, and converting the search data into a language vector into a search language vector. a feature amount conversion step of converting the search language vector into a search data feature using a feature conversion model obtained by machine learning the relationship between the model language vector and the model feature; a feature comparison step of comparing the search data feature with each of the model features; and a selection of selecting at least one trained model from the plurality of trained models based on the comparison result of the feature comparison step. and a step.

According to a preferred embodiment, in the selection step, a trained model corresponding to a model feature having the highest degree of similarity to the search data feature is selected.

According to a preferred embodiment, the learned model selection method includes a reception step of accepting the analysis target or language data corresponding to the analysis target as search data; a conversion step of converting the search data into data in the same format as the teacher data, a comparison step of comparing the search data with the teacher data, and at least one of the plurality of trained models based on the comparison result of the comparison step. a selection step of selecting a trained model.

According to a preferred embodiment, the teacher data constitutes a plurality of teacher data sets respectively corresponding to the plurality of trained models, and in the comparison step, the plurality of teacher data sets are sequentially selected and selected. The teacher data of the selected teacher data set and the search data are compared, and the degree of similarity between the teacher data and the search data is calculated for each of the teacher data sets, and in the selection step, the degree of similarity is used to select the plurality of At least one teacher data set is selected from the teacher data sets, and a trained model corresponding to the selected teacher data set is selected.

According to a preferred embodiment, the early search data is an image.

A trained model selection device according to the present invention is a trained model selection device that selects a trained model for analyzing an analysis target from a plurality of trained models, and wherein a trained model is assigned to each of the plurality of trained models. a model language vector that is a language vector generated from a language label, a model feature that is a feature of each of the plurality of trained models, or a teacher used for machine learning of each of the plurality of trained models. A trained model corresponding to the analysis target is selected using the degree of similarity between the data and the analysis target.

The trained model selection program according to the present invention is a trained model selection program that selects a trained model for analyzing an analysis target from a plurality of trained models, and is a trained model selection program that selects a trained model for analyzing an analysis target from a plurality of trained models, and is a trained model selection program that corresponds to each of the plurality of trained models. A model language vector that is a language vector, a model feature that is a feature of each of the plurality of trained models, or a link between the training data used for machine learning of each of the plurality of trained models and the analysis target. Using the degree of similarity, a computer is caused to execute a process of selecting a trained model according to the analysis target.

According to the present invention, a trained model suitable for an analysis target can be quickly selected from a plurality of trained models.

1 is a block diagram of a learned model selection device according to Embodiment 1 of the present invention. FIG. This is an example of a vector space. 3 is a flowchart showing a processing procedure of a learned model selection method according to Embodiment 1 of the present invention. This is an example of language data for search. 7 is a flowchart showing a processing procedure of a learned model selection method according to a modification of Embodiment 1 of the present invention. FIG. 2 is an explanatory diagram showing an example of generation of a model language vector. FIG. 2 is a block diagram of a learned model selection device according to a second embodiment of the present invention. This is an example of a vector space. 7 is a flowchart showing a processing procedure of a learned model selection method according to Embodiment 2 of the present invention. FIG. 3 is a block diagram of a learned model selection device according to Embodiment 3 of the present invention. 12 is a flowchart showing a processing procedure of a learned model selection method according to Embodiment 3 of the present invention.

Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. Note that the present invention is not limited to the embodiments described below.

[Embodiment 1]
Embodiment 1 of the present invention will be described below.

(Learned model selection device)
FIG. 1 is a block diagram of a learned model selection device 1 according to the first embodiment. The trained model selection device 1 has a function of selecting a trained model for analyzing an analysis target from a plurality of trained models. In this embodiment, the analysis target is an image and the analysis method is segmentation, but the analysis target and analysis method are not particularly limited.

The trained model selection device 1 can be configured with a general-purpose computer, and its hardware configuration includes a processor such as a CPU or GPU, a main storage device such as a DRAM or SRAM (not shown), and an HDD or SSD. It is equipped with an auxiliary storage device 10. The auxiliary storage device 10 includes a plurality of learned models M1 to Mn machine-learned using images, model language vectors V1 to Vn, as well as a learned model selection device 1 such as a learned model selection program. Contains various programs.

The learned models M1 to Mn are artificial intelligence models that have been machine learned using different learning data sets. Note that each learning data set may include the same teacher image in common, but it is preferable that each learning data set is composed of teacher images having different characteristics. For example, trained model M1 is machine learned using a training dataset that includes many teacher images of round particles, and trained model M2 is machine learned using a training dataset that includes many teacher images of cherries. ing. The number of trained models M1 to Mn is n (n≧2), but the number is not particularly limited.

In this embodiment, the model language vectors V1 to Vn are generated from the language labels assigned to each of the trained models M1 to Mn. For example, the trained model M2 has been machine learned using a group of images that includes many images of cherries, and can analyze images of cherries and objects similar to cherries with high accuracy. Color, coordinates" = "Round, small, red, whatever" is given a language label. The language label may be assigned by a human or by an algorithm such as image caption, mirror GAN, image to text, etc. By converting this language label into a language vector, a model language vector V2 in which (shape, size, color, coordinates) is encoded is generated.

Note that in the above example, the number of dimensions of the model language vector is four, but the number of dimensions is not particularly limited, and may be several hundred dimensions, for example. The k-dimensional model language vector can be represented by a vector space V shown in FIG. 2, for example. Further, in the above example, the model language vector included only elements that humans can recognize, but it may also include elements that humans cannot recognize.

The trained model selection device 1 includes a reception section 11, a language vector conversion section 12, a language vector comparison section 13, a selection section 14, and an analysis section 15 as functional blocks. In this embodiment, each of these units is realized in software by the processor of the learned model selection device 1 reading out the learned model selection program into the main storage device and executing it.

(Trained model selection method)
The functions of the above-mentioned parts of the trained model selection device 1 will be explained based on FIG. 3. FIG. 3 is a flowchart showing the processing procedure of the learned model selection method according to the present embodiment. These steps S1 to S8 are executed by the trained model selection device 1. Note that from the viewpoint of the final purpose, steps S1 to S8 are processing steps of the image analysis method, and are divided into steps S1 to S3, which are preprocessing steps, and steps S4 to S8, which are analysis steps. Note that the calculation load can be reduced by creating a database for the preprocessing process.

In step S1, the trained models M1 to Mn are stored in the auxiliary storage device 10. Note that the learned models M1 to Mn may be stored in an external storage device or cloud.

In step S2, the language labels assigned to each of the trained models M1 to Mn are converted into model language vectors V1 to Vn. Step S2 may be executed by the language vector converter 12.

In step S3, the model language vectors V1 to Vn are stored in the auxiliary storage device 10. Note that the model language vectors V1 to Vn may be stored in an external storage device or cloud.

In step S4 (reception step), the reception unit 11 receives the analysis target or the language data corresponding to the analysis target as search data. In this embodiment, the reception unit 11 receives language data corresponding to the analysis target as search data. As for the linguistic data, for example, when an image of a cherry is to be analyzed, the user performs a search by inputting the linguistic data "small round red particles" into the search field R, as shown in FIG. The language data is not particularly limited as long as it expresses the content or characteristics of the analysis target, and may be the name of the analysis target (for example, "cherry").

Additionally, the search data may be the analysis target itself. When the analysis data is an image, the user performs a search by uploading the image data to the learned model selection device 1. Uploaded images are converted into linguistic data using algorithms such as image captions, mirror GAN, and image to text.

In step S5 (language vector conversion step), the language vector conversion unit 12 converts the search data into a language vector and converts it into a search language vector Va. For example, if the search data is "round and small red particles", the language vector conversion unit 12 converts the search language vector Va such that "shape, size, color, coordinates" = "round, small, red, whatever". Convert to The method of converting search data into a search language vector is not particularly limited, and a dictionary that vectorizes words or an artificial intelligence model that has been machine learned to vectorize language data may be used. Note that the search language vector corresponding to the search word may have a variable length.

In step S6 (language vector comparison step), the language vector comparison unit 13 compares the search language vector Va converted in step S5 with each of the model language vectors V1 to Vn. In this embodiment, the language vector comparison unit 13 calculates the similarity of the search language vector Va to each of the model language vectors V1 to Vn by numerical calculation. The degree of similarity can be determined by a method using an algorithm such as cosine similarity or pattern matching, or by a known technique of inference using a trained model that has learned a data set in which the degree of similarity has been evaluated subjectively by a person.

In step S7 (selection step), the selection unit 14 selects at least one learned model from the learned models M1 to Mn based on the comparison result in step S6. In the present embodiment, the selection unit 14 selects the trained model corresponding to the model language vector having the highest degree of similarity to the search language vector Va from among the model language vectors V1 to Vn as the trained model suitable for the analysis target. do. Note that the selection unit 14 may select a plurality of trained models as long as they are trained models that correspond to model language vectors that have a high degree of similarity to the search language vector Va. Alternatively, trained models may be ranked in descending order of similarity to the search language vector Va, and the user may select one of the trained models.

In step S8, the analysis unit 15 analyzes the analysis target using the trained model selected in step S7.

(Modified example)
FIG. 5 is a flowchart showing the processing procedure of a learned model selection method according to a modification of the first embodiment, and these processes are executed by the learned model selection device 1 shown in FIG. 1. This modification is characterized in that the model language vectors V1 to Vn stored in the auxiliary storage device 10 are generated from the language labels given to the teacher data used for machine learning of each of the trained models M1 to Mn. , is different from the above. That is, the flowchart shown in FIG. 5 differs from the flowchart shown in FIG. 3 in that step S2 is replaced with step S2'. Therefore, the explanation of steps S1, S3 to S8 will be omitted below.

In step S2', the language labels given to the teacher data used for machine learning of each of the trained models M1 to Mn are converted into model language vectors V1 to Vn. For example, as shown in FIG. 6, each of the training data T1 to Tm for generating the learned model Mx (2≦x≦n) is given a linguistic label such as “round” or “square”. There is. The language label may be assigned by a human or by an algorithm such as image caption, mirror GAN, image to text, etc. A model language vector Vx is generated by converting these language labels into language vectors.

Although the specific method of language vectorization is not particularly limited, examples include the following methods.
・One-hot vector: A method of storing words as individual vectors ・Distributed representation: A method of representing words as low-dimensional vectors ・Bag-of-words: A method of using the number of times a word appears in a sentence as an element of a vector For methods, see, for example, https://deepage.net/bigdata/machine_learning/2016/09/02/word2vec_power_of_word_vector.html, https://deepage.net/machine_learning/2017/01/08/doc2vec.html I want to be

Additionally, although the above method vectorizes words, if the text is generated using image to text or GAN (Generative Adversarial Networks), it is also possible to vectorize the text (e.g., Doc2Vec, BERT). .

(Brief Summary)
As described above, in the first embodiment, a trained model corresponding to the analysis target is selected from the trained models M1 to Mn by using the model language vectors V1 to Vn corresponding to each of the trained models M1 to Mn. ing. Since language vectors require less calculation to compare vectors, a trained model suitable for an analysis target can be selected more quickly than in the prior art.

[Embodiment 2]
Embodiment 2 of the present invention will be described below. In Embodiment 2, members having the same functions as those in Embodiment 1 described above are designated by the same reference numerals, and detailed description thereof will be omitted.

(Learned model selection device)
FIG. 7 is a block diagram of a trained model selection device 1' according to the second embodiment. Like the trained model selection device 1 shown in FIG. 1, the trained model selection device 1' has a function of selecting a trained model for analyzing an analysis target from a plurality of trained models. In this embodiment as well, the analysis target is an image and the analysis method is segmentation, but the analysis target and the analysis method are not particularly limited.

The hardware configuration of the trained model selection device 1' is the same as that of the trained model selection device 1. The auxiliary storage device 10 of the trained model selection device 1' includes n trained models M1 to Mn, model features F1 to Fn, feature conversion models MC, and trained model selection programs. Various programs for operating the model selection device 1' are stored.

The model feature quantities F1 to Fn are the feature quantities of each of the learned models M1 to Mn, respectively. In this embodiment, the features of the trained model are latent variables that accompany the trained model. For example, if the trained model is a neural network, the hyperparameters and feature filters used to generate the trained model It is. The model feature amount can be represented by a vector space W shown in FIG. 8, for example.

The feature quantity conversion model MC is an artificial intelligence model in which the relationship (probability distribution) between the model language vectors V1 to Vn (FIG. 1) and the model feature quantities F1 to Fn is machine-learned in advance. The machine learning method is not particularly limited, but for example, linear regression or random forest can be applied.

The trained model selection device 1' includes a reception section 11, a language vector conversion section 12, a selection section 14, an analysis section 15, a feature amount conversion section 16, and a feature amount comparison section 17 as functional blocks. ing. That is, the trained model selection device 1' is the trained model selection device 1 shown in FIG. In this embodiment, each of these parts is realized in software by the processor of the learned model selection device 1' reading out the learned model selection program into the main storage and executing it.

(Trained model selection method)
The functions of the above-mentioned parts of the trained model selection device 1' will be explained based on FIG. 9. FIG. 9 is a flowchart showing the processing procedure of the trained model selection method according to the present embodiment, and these steps S11 to S21 are executed by the trained model selection device 1'. Note that from the viewpoint of the final purpose, steps S11 to S21 are processing steps of the image analysis method, and are divided into steps S11 to S15, which are preprocessing steps, and steps S16 to S21, which are analysis steps. Note that the calculation load can be reduced by creating a database for the preprocessing process.

In step S11, the trained models M1 to Mn are stored in the auxiliary storage device 10. Step S11 is similar to step S1 shown in FIG.

In step S12, the language labels assigned to each of the learned models M1 to Mn are converted into model language vectors V1 to Vn. Step S12 is similar to step S2 shown in FIG. 3, but may be similar to step S2' shown in FIG.

In step S13, model feature quantities F1 to Fn of learned models M1 to Mn are calculated.

In step S14, machine learning is performed on the relationship between the model language vectors V1 to Vn generated in step S12 and the model feature quantities F1 to Fn generated in step S13. Thereby, when a language vector is input, a feature amount conversion model MC is generated that outputs a feature amount corresponding to the language vector.

In step S15, the model feature quantities F1 to Fn and the machine-learned feature quantity conversion model MC are stored in the auxiliary storage device 10. Note that these may be saved in an external storage device or cloud.

In step S16 (reception step), the reception unit 11 receives the analysis target or the language data corresponding to the analysis target as search data. Step S16 is similar to step S4 shown in FIG.

In step S17 (language vector conversion step), the language vector conversion unit 12 converts the search data into a language vector and converts it into a search language vector Va. Step S17 is similar to step S5 shown in FIG.

In step S18 (feature amount conversion step), the feature amount conversion unit 16 converts the search language vector Va converted in step S17 into the search data feature amount Fa using the feature amount conversion model MC. The feature conversion model MC performs machine learning on the relationship between the model language vectors V1 to Vn of the trained models M1 to Mn and the model features F1 to Fn, so when the search language vector Va is input, the search A search data feature amount Fa, which is a feature amount corresponding to the language vector Va, is output.

In step S19 (feature amount comparison step), the feature amount comparison unit 17 compares the search data feature amount Fa converted in step S18 with each of the model feature amounts F1 to Fn. In this embodiment, the feature comparison unit 17 calculates the degree of similarity of the search data feature Fa to each of the model features F1 to Fn by numerical calculation. The degree of similarity can be determined by a method using an algorithm such as cosine similarity or pattern matching, or by a known technique of inference using a trained model that has learned a data set in which the degree of similarity has been evaluated subjectively by a person.

In step S20 (selection step), the selection unit 14 selects at least one learned model from the learned models M1 to Mn based on the comparison result in step S19. In the present embodiment, the selection unit 14 selects a trained model corresponding to a model feature having the highest degree of similarity to the search data feature Fa among the model features F1 to Fn as a trained model suitable for the analysis target. select. Note that the selection unit 14 may select a plurality of trained models as long as they correspond to model features having a high degree of similarity to the search data feature Fa.

In step S21, the analysis unit 15 analyzes the analysis target using the learned model selected in step S20.

(Brief Summary)
As described above, in the second embodiment, by using the model feature quantities F1 to Fn, which are the feature quantities of each of the trained models M1 to Mn, a trained model according to the analysis target is generated from the trained models M1 to Mn. Selected. Similar to language vectors, feature quantities require a small amount of calculation to compare feature quantities with each other, so compared to conventional techniques, a trained model suitable for an analysis target can be selected more quickly.

[Embodiment 3]
Embodiment 3 of the present invention will be described below. In Embodiment 3, members having the same functions as those in Embodiments 1 and 2 described above are denoted by the same reference numerals, and detailed description thereof will be omitted.

(Learned model selection device)
FIG. 10 is a block diagram of a trained model selection device 1'' according to the third embodiment.The trained model selection device 1'' is similar to the trained model selection devices 1 and 1' shown in FIGS. 1 and 7. , has a function of selecting a trained model for analyzing an analysis target from a plurality of trained models. In this embodiment, the analysis target and teacher data are images, and the analysis method is segmentation, but the data formats and analysis method of the analysis target and teacher data are not particularly limited.

The hardware configuration of the learned model selection device 1'' is the same as that of the learned

model selection devices

1 and 1'.The auxiliary storage device 10 of the learned model selection device 1'' stores n learned models M1 to M1. In addition to Mn and n teacher data sets S1 to Sn, various programs for operating the learned model selection device 1'', such as a learned model selection program, are stored.

The teacher data sets S1 to Sn correspond to learned models M1 to Mn, respectively, and the teacher data used for machine learning of the learned models M1 to Mn constitute the teacher data sets S1 to Sn. That is, the teacher data set Sk (1≦k≦n) is composed of m pieces of teacher data k-1 to km (m is an undefined integer) used for machine learning of the trained model Mk. . Note that in recent machine learning, transfer learning is often performed, so m is approximately several hundred.

The learned model selection device 1'' includes a reception section 11, a selection section 14, an analysis section 15, a search data conversion section 18, and a search data comparison section 19 as functional blocks. The model selection device 1'' is the learned model selection device 1 shown in FIG. 1 in which the language vector comparison section 13 is replaced with a search data conversion section 18 and a search data comparison section 19. In this embodiment, each of these parts is realized in software by the processor of the learned model selection device 1'' reading out the learned model selection program into the main storage and executing it.

(Trained model selection method)
The functions of the above-mentioned parts of the learned model selection device 1'' will be explained based on FIG. 11. FIG. 11 is a flowchart showing the processing procedure of the learned model selection method according to the present embodiment, and these steps S31 to S37 is executed by the trained model selection device 1''. Note that from the viewpoint of the final purpose, steps S31 to S37 are processing steps of the image analysis method, and are divided into step S31, which is a preprocessing step, and steps S32 to S37, which are analysis steps. Note that the calculation load can be reduced by creating a database for the preprocessing process.

In step S31, the trained models M1 to Mn and the teacher data sets S1 to Sn are stored in the auxiliary storage device 10.

In step S32 (reception step), the reception unit 11 receives the analysis target or the language data corresponding to the analysis target as search data.

If the search data is in the same format as the teacher data (image in this embodiment) (YES in step S33), the process moves to step S35. If the search data is in a format different from the teacher data (for example, language data) (NO in step S33), the process moves to step S34.

In step S34 (conversion step), the search data conversion unit 18 converts the search data into a teacher data format (image). An algorithm such as mirror GAN can be used for conversion to an image.

In step S35 (comparison step), the search data comparison unit 19 compares the search data (image) with the teacher data. Specifically, the search data comparison unit 19 sequentially selects the teacher data sets S1 to Sn, compares each teacher data of the selected teacher data sets with the search data, and compares the teacher data and the search data for each teacher data set. Calculate the degree of similarity with the data. In this embodiment, the search data comparison unit 19 calculates the average value or maximum value of similarity. The degree of similarity between the training data and the search data can be determined by known techniques such as algorithmic methods such as cosine similarity and pattern matching, and inference using a trained model trained on a data set that evaluates the degree of similarity based on human subjectivity. can. Note that the search data comparison unit 19 does not need to compare all the teacher data of each teacher data set.

In step S36 (selection step), the selection unit 14 selects at least one learned model from the learned models M1 to Mn based on the comparison result in step S35. In the present embodiment, the selection unit 14 selects the trained model corresponding to the teaching data set with the largest average or maximum similarity between the teaching data and the search data as the trained model suitable for the analysis target. . Note that the selection unit 14 may select a plurality of trained models as long as they are trained models that correspond to a teacher data set with a large average value or maximum value of similarity between each teacher data and the search data.

In step S37, the analysis unit 15 analyzes the analysis target using the trained model selected in step S36.

(Brief Summary)
As described above, in the third embodiment, by using the degree of similarity between the training data used for machine learning of each of the trained models M1 to Mn and the analysis target, A trained model is selected. In this embodiment, since there is no need to input test data to a trained model, a trained model suitable for an analysis target can be selected more quickly than in the prior art.

Note that in this embodiment, when the analysis target is an image and the search data is language data, the language data is converted to an image and compared with the teacher data, but the present invention is not limited to this. If the analysis target is in a format other than an image and the search data is language data, the language data is converted to data in the same format as the teacher data and compared with the teacher data.

(Additional notes)
Although the embodiments of the present invention have been described above, the present invention is not limited to the above embodiments, and various changes can be made without departing from the spirit thereof.

1 Trained model selection device 1' Trained model selection device 1'' Trained model selection device 10 Auxiliary storage device 11 Reception unit 12 Language vector conversion unit 13 Language vector comparison unit 14 Selection unit 15 Analysis unit 16 Feature amount conversion unit 17 Features Quantity comparison section 18 Search data conversion section 19 Search data comparison section F1 to Fn Model feature quantity Fa Search data feature quantity M1 to Mn Learned model MC Feature quantity conversion model S1 to Sn Teacher data set T1 to Tm Teacher data T1-1 ～T1-m Teacher data Tn-1～Tn-m Teacher data V1～Vn Model language vector Va Search language vector

Claims

A trained model selection method for selecting a trained model for analyzing an analysis target from a plurality of trained models, the method comprising:
A model language vector that is a language vector corresponding to each of the plurality of trained models, a model feature quantity that is a feature quantity of each of the plurality of trained models, or a machine learning method for each of the plurality of trained models. A trained model selection method that selects a trained model according to the analysis target using a degree of similarity between used teacher data and the analysis target.
The trained model selection method according to claim 1, wherein the model language vector is generated from a language label given to each of the trained models.
The learned model selection method according to claim 1, wherein the model language vector is generated from a language label given to teacher data used for machine learning of each of the learned models.
a reception step of accepting the analysis target or language data corresponding to the analysis target as search data;
a language vector conversion step of converting the search data into a language vector and converting it into a search language vector;
a language vector comparison step of comparing the search language vector with each of the model language vectors;
a selection step of selecting at least one trained model from the plurality of trained models based on the comparison result of the language vector comparison step;
The learned model selection method according to claim 2 or 3, comprising:
5. The trained model selection method according to claim 4, wherein in the selection step, a trained model corresponding to a model language vector having the highest degree of similarity to the search language vector is selected.
a reception step of accepting the analysis target or language data corresponding to the analysis target as search data;
a language vector conversion step of converting the search data into a language vector and converting it into a search language vector;
a feature quantity conversion step of converting the search language vector into a search data feature quantity using a feature quantity conversion model obtained by machine learning the relationship between the model language vector and the model feature quantity;
a feature comparison step of comparing the search data feature with each of the model feature;
a selection step of selecting at least one trained model from the plurality of trained models based on the comparison result of the feature value comparison step;
The learned model selection method according to claim 1, comprising:
The trained model selection method according to claim 6, wherein in the selection step, a trained model corresponding to a model feature having the highest degree of similarity to the search data feature is selected.
a reception step of accepting the analysis target or language data corresponding to the analysis target as search data;
When the search data is linguistic data, a conversion step of converting the linguistic data into data in the same format as the teacher data;
a comparison step of comparing the search data with the teacher data;
a selection step of selecting at least one trained model from the plurality of trained models based on the comparison result of the comparison step;
The learned model selection method according to claim 1, comprising:
The training data constitutes a plurality of training data sets respectively corresponding to the plurality of trained models,
In the comparison step, the plurality of teacher data sets are sequentially selected, the teacher data of the selected teacher data set and the search data are compared, and the degree of similarity between the teacher data and the search data is determined for each of the teacher data sets. Calculate,
9. The method according to claim 8, wherein in the selection step, at least one teacher data set is selected from the plurality of teacher data sets using the similarity, and a trained model corresponding to the selected teacher data set is selected. Trained model selection method.
The trained model selection method according to claim 9, wherein the first-term search data is an image.
A trained model selection device that selects a trained model for analyzing an analysis target from a plurality of trained models,
A model language vector that is a language vector generated from a language label assigned to each of the plurality of trained models, a model feature that is a feature of each of the plurality of trained models, or a model feature that is a feature of each of the plurality of trained models, or a plurality of trained models. A trained model selection device that selects a trained model according to the analysis target using a degree of similarity between training data used for machine learning of each model and the analysis target.
A trained model selection program that selects a trained model for analyzing an analysis target from multiple trained models,
A model language vector that is a language vector corresponding to each of the plurality of trained models, a model feature quantity that is a feature quantity of each of the plurality of trained models, or a machine learning method for each of the plurality of trained models. A trained model selection program that causes a computer to execute a process of selecting a trained model corresponding to the analysis target using a degree of similarity between used teaching data and the analysis target.