JP2022191637A - AI platform system - Google Patents

Abstract

To realize environment in which a user who has no expert knowledge can make and try an AI model meeting respective needs.SOLUTION: A system stores AI selection tag information including a business type, a use purpose and a use scene, and preset information constituted by a hyper parameter group including a plurality of hyper parameters different from each other for each AI model. The system extracts a plurality of corresponding AI models as learning object AI models by using the business type, the use purpose and the use scene selected by an AI usage user. Learning processing executes the learning processing by the number of the hyper parameters, performs first learning processing which generates a plurality of learned AI models for each hyper parameter, and executes the first learning processing to all of the plurality of learning object AI models. The learning processing extracts a learned AI model from among the plurality of generated learned AI models for each hyper parameter, and provides accuracy in learning of each extracted learned AI model, as a learning result of each of the plurality of learning object AI models.SELECTED DRAWING: Figure 1

Description

The present invention relates to AI model utilization/service technology.

In recent years, AI models have been actively utilized and introduced. An AI model can be generated by an AI engine learning with a predetermined AI algorithm. However, as described in Patent Document 1, what kind of AI algorithm is applied to the AI engine (module for performing machine learning, deep learning, etc.) that serves as a container for learning requires specialized knowledge. and

In Patent Document 1, a user selects an AI engine and an algorithm, and applies multiple patterns of hyperparameters, which are tuning parameters, to generate an AI model. As a result, even a user without specialized knowledge can use an AI model to which highly accurate hyperparameters are applied. It also proposes to compare between multiple AI engines and algorithms belonging to the same data analysis method.

JP 2020-52514 A Patent No. 6703264

However, Patent Document 1 has a problem that it is difficult for a user to select an AI engine and an algorithm as a premise. There are countless combinations of AI engines and algorithms, and with the rapid development of AI-related technologies in recent years, new algorithms are appearing every day. For this reason, it is not easy for a user to grasp the characteristics of AI engines and algorithms and select a combination of AI engines and algorithms.

Therefore, even if there is a mechanism for simply searching for hyperparameters that increase accuracy for the combination of the selected AI engine and algorithm (for example, Patent Documents 1 and 2), it is not possible to easily introduce an AI model. is the actual situation. In addition, Patent Document 1 proposes comparison between multiple AI engines and algorithms belonging to the same data analysis method, but this is also the same as the selection itself of the AI engine and algorithm selected for comparison. is not easy either.

In the conventional system, the entrance part of "AI selection" was a high hurdle, including users who had no knowledge at all, and it was a factor that made it difficult to introduce and utilize AI models.

In addition, as mentioned above, US Pat. No. 5,700,000 also proposes comparing between multiple AI engines and algorithms belonging to the same data analysis technique. However, in the mechanism of Patent Document 1, it is necessary to select a combination of AI engines and algorithms to create an AI model, and select a combination of AI engines and algorithms again to create an AI model. Therefore, each AI model must be generated individually, which is very time-consuming.

In order to promote the introduction and utilization of AI models, it is extremely important to have a platform environment from the perspective of "trying to create" AI models, but in the past, system engineers, etc. had a certain level of technical knowledge. It was limited to a mechanism that allows users to create AI models without specialized knowledge and to search for optimal tuning parameters.

In particular, with the rapid development of AI-related technologies, the needs of users who use AI models have become diverse, and both training data and inference data applied to AI models can change in a short time span. . If the data changes, the AI model once generated cannot be used as it is, so it is necessary to re-evaluate the AI model and generate a new AI model. However, conventionally, there was no mechanism for centrally managing the environment for re-evaluating the currently used AI model and generating a new AI model, so the management burden on the user was heavy.

Accordingly, it is an object of the present invention to provide an AI platform system that realizes an environment in which users without specialized knowledge can create and test AI models that meet their needs.

(1) The present invention is an AI platform system that provides a platform for AI users to create and test AI models that meet their needs. This system includes a first storage unit that stores AI selection tag information associated with each AI model provided by an AI vendor, including industry types, applications, and usage scenes, and preset information for learning AI models provided by the AI vendor. and a second storage unit that stores hyperparameter preset information that is generated for each AI model and configured by a hyperparameter group including a plurality of hyperparameters that are different from each other.

In addition, the AI user controls to select the industry, application and usage scene through the user terminal, searches for AI selection tag information using the selected industry, application and usage scene, and selects the corresponding AI model as the learning target AI An AI model proposal unit that extracts multiple models as a model, a learning processing unit that generates a trained AI model by executing the learning process of the learning target AI model using the learning data specified by the AI user, and the AI user and an inference processing unit that executes inference processing of the trained AI model using designated inference data to generate an inference result.

Then, the learning processing unit sets each hyperparameter of the hyperparameter group linked to the learning target AI model as a learning parameter, executes the learning process by the number of hyperparameters, and creates a plurality of learned AI models for each hyperparameter. The first learning process to be generated is performed, and the first learning process is performed on all of the plurality of extracted learning target AI models. The AI model proposal unit selects a trained AI model from among a plurality of hyperparameter-specific trained AI models generated based on the learning accuracy generated along with the first learning process for each of the plurality of learning target AI models. The accuracy at the time of learning of each of the extracted learned AI models is provided to the AI user as the learning results of each of the plurality of learning target AI models.

(2) In (1) above, the AI model proposal unit causes the user terminal to display a selection screen including selection areas for each type of industry, application, and usage scene based on the AI selection tag information, and displays the selection screen through the selection screen. It is possible to provide a user terminal with an upload screen for accepting learning data to be applied to a plurality of learning target AI models extracted based on each selection information of industry type, application, and usage scene from an AI user. The learning processing unit can be configured to use the learning data designated through the upload screen and execute the learning process for the plurality of learning target AI models extracted through the selection screen.

(3) In (1) or (2) above, the AI model proposing unit generates a learning result screen in which the learning results of each of the plurality of learning target AI models are ranked in descending order of accuracy during learning, and generates a learning result screen for the user terminal. can be configured to provide

(4) In (1) to (3) above, the function classification information based on the characteristics of the AI algorithm can be configured to be associated with each AI model provided by the AI vendor. At this time, the AI model proposal unit can provide the user terminal with a function selection screen based on the function classification information, and extracts an AI model linked to the function classification selected on the function selection screen as a learning target AI model. can be configured to

(5) In (1) to (4) above, further include an external service cooperation unit that accepts annotation requests for learning data from AI users and provides annotation requests to support providers that implement annotations. Can be configured. The external service linking part provides the learning data received from the AI user to the support business operator, receives the annotated learning data that has been annotated from the support business business operator, and stores the learning data assigned to each user. Can be configured to save to a region.

(6) The program of the present invention is a program executed by a computer device that provides a platform for AI users to create and test AI models that meet their needs, and makes the computer device realize the following functions.

A first function for storing AI selection tag information associated with each AI model provided by an AI vendor, including industry types, applications, and usage scenes; A second function that stores hyperparameter preset information that is generated every time and is composed of a group of hyperparameters that include multiple hyperparameters that are different from each other. , using the selected industry, application and usage scene to search for AI selection tag information, extracting multiple applicable AI models as learning target AI models, using learning data specified by the AI user, Using the fourth function that executes the learning process of the AI model to be learned and generates the trained AI model, and the inference data specified by the AI user, execute the inference process of the trained AI model and generate the inference result. 5th function.

Then, the fourth function sets each hyperparameter of the hyperparameter group linked to the learning target AI model as a learning parameter, executes the learning process by the number of hyperparameters, and creates a plurality of trained AI models for each hyperparameter. While performing the first learning process to generate, the first learning process is performed for all of the plurality of the extracted learning target AI models, and the third function performs the first learning process for each of the plurality of learning target AI models A trained AI model is extracted from a plurality of trained AI models by hyperparameter generated based on the learning accuracy generated along with Provide AI users with learning results for each AI model to be learned.

Since the present invention extracts an AI model that matches the AI user from the type of industry, role, and usage scene, the AI model that matches the AI user is automatically presented even if the user does not have specialized knowledge. .

Then, for one learning target AI model, a plurality of trained AI models with different hyperparameters (trained AI models by hyperparameters) are generated at once, and among them, a trained AI model is generated based on the accuracy at the time of learning. to extract This is performed for each of the plurality of learning target AI models presented to the AI user, and collective learning results of the plurality of learning target AI models are provided. As a result, it is possible to reduce the time and cost required for searching for hyperparameters, etc., and realize an environment in which a plurality of AI models compete with each other.

According to the present invention, even if an AI user does not have specialized knowledge, an AI model that matches the user is automatically presented, and a plurality of AI models are competed to meet various needs. can make it easier to propose Therefore, it is possible to realize an environment in which users without specialized knowledge can create and test AI models that meet their needs.

1 is a network configuration diagram of the AI platform system of the first embodiment; FIG. 1 is a configuration block diagram of an AI platform device according to a first embodiment; FIG. It is an example of the AI model proposal screen of 1st Embodiment. 4 is a diagram showing an example of AI selection tag information according to the first embodiment; FIG. FIG. 10 is a diagram showing an example of “industry” of AI selection tag information; FIG. 5 is a diagram showing an example of “role (use category)” of AI selection tag information; FIG. 4 is a diagram showing an example of AI algorithm attribute information of AI selection tag information; FIG. 11 is a diagram showing an example of a screen of a proposal UI “start from industry/role setting” using the AI selection tag information of the first embodiment; It is a figure which shows an example of the registration screen of the data for learning of 1st Embodiment. It is a figure which shows an example of the screen which displays AI model under preparation of 1st Embodiment. It is an example of the screen which displays the learning result of the AI model produced by 1st Embodiment. FIG. 10 is a diagram showing an example of a detail (Detail) screen of AI model learning results; FIG. 11 is a diagram showing an example of a screen of a proposal UI “start from selection of desired function” according to the first embodiment; FIG. 11 is a diagram showing an example of an inference data designation screen in the UI “Let's use an AI model (inference)” of the first embodiment; It is an example of the screen which displays the state in inference of 1st Embodiment. It is a figure which shows an example of the screen which displays the inference result of 1st Embodiment. FIG. 11 is a diagram showing an example of a detail screen of an inference result; It is an example of the screen which displayed the inference result of the time-series related AI model. FIG. 4 is a diagram showing an example of a project management screen according to the first embodiment; FIG. It is an example of a screen displaying an AI model used in each project on the project management screen of the first embodiment. It is a figure which shows an example of the Data management screen of 1st Embodiment. It is a figure which shows an example of the AI model management screen of 1st Embodiment. It is an example of a screen displayed when downloading an AI model on the AI model management screen. FIG. 10 is a diagram showing a display example of an AI model being created on an AI model management screen; It is a figure which shows an example of the option service management screen of 1st Embodiment. FIG. 4 is a diagram for explaining a mechanism for "competing" a plurality of AI models according to the first embodiment; FIG. 4 is a diagram for explaining a cooperation interface function of the first embodiment; FIG. It is a figure which shows the system flow of 1st Embodiment. It is a figure which shows the system flow of 1st Embodiment. It is a figure which shows the system flow of 1st Embodiment.

(First embodiment)
1 to 30 are diagrams for explaining the first embodiment. FIG. 1 is a network configuration diagram of an AI platform system provided by an AI platform device (hereinafter referred to as a management device) 100. As shown in FIG.

This system realizes a platform environment in which an AI user (user terminal 300), an AI vendor (vendor terminal 500), and a support provider (support provider terminal 600) are connected with a management device 100 at the center.

The platform provided by the management device 100 includes each environment of "search", "create", and "try", an AI information portal site as the "search" environment, and an AI development platform as the "create" and "try" environment. We will provide a comprehensive AI platform to users who want to create and introduce AI (users who want to use AI, hereinafter referred to as AI users). As for the functions of the AI information portal site of the "search" environment, the technology of publicly known information portal sites can be applied, so a detailed explanation will be omitted. The explanation will focus on the AI development platform.

Here, the terms used in this embodiment will be explained. Taking the machine learning engine as an example, various algorithms such as decision tree, random forest, neural network, logistic regression, naive Bayes classifier, support vector machine, etc. are applied to the machine learning engine according to its application. In other words, an AI model is a combination of an AI engine and an algorithm, such as a "neural network-based machine learning engine." In addition to machine learning, AI models such as deep learning with a deep neural network hierarchy and other known AI models are also targets of this system, and are not limited to the AI models exemplified in this embodiment.

Next, the learned AI model is the result of applying the learning data to the pre-learning AI model and making it learn, for example, it is possible to identify, classify, and analyze data patterns and objects in the data. It is a program module that can A trained AI model mainly includes a program part and learning parameters (hyperparameters).

AI models are combinations of AI engines and algorithms as described above, and have various characteristics. For example, identification systems include speech recognition, image recognition, moving image recognition, and language analysis, and prediction systems include numerical prediction, matching, intention prediction, and needs prediction. Then, an AI model that combines identification and prediction is constructed, and an AI model with various characteristics such as object detection, character recognition, face recognition, and automatic driving is provided.

AI models can include aspects in which multiple algorithms are applied to a single AI model. In addition to the individual models of the discrimination system and the prediction system, it is possible to include a mode of a composite model in which individual models of the discrimination system and the prediction system are connected and combined. A composite model may include two or more discrimination systems or two or more prediction systems, and may be a model group in which a plurality of single models are connected. Note that the composite model can be understood in units of algorithms. In other words, in addition to a model group in which a plurality of AI models are connected, a single AI model composed of a plurality of algorithms can also be regarded as a composite model.

<System configuration>
FIG. 2 is a configuration block diagram of the management device 100. As shown in FIG. The management device 100 includes a communication device 110 , control devices 120 , 130 and 140 and storage devices 150 , 160 and 170 . The communication device 110 performs connection control and data communication control between the user terminal 300 , the vendor terminal 500 and the support provider terminal 600 connected to the management device 100 .

The control device 120 mainly performs registration management of AI models provided by AI vendors, management of preset information for learning processing, learning processing, and inference (prediction) processing. The AI registration unit 121 performs processing of vendor registration and AI model registration with the vendor terminal 500 . The AI registration unit 121 stores vendor information input through a vendor registration screen (not shown) in a predetermined storage area, and stores an AI model uploaded from the vendor terminal 500 in the registered AI model 151 . At this time, the AI model uploaded from the vendor terminal 500 includes both or one of the pre-learning AI model and the learned AI model. If the data volume of the program module of the AI model is large, instead of uploading from the vendor terminal 500 through the network, an electronic medium (HDD, flash memory, etc.) storing the AI model is sent to the operator of the management device 100. It can also be configured such that it is physically sent and the operating company registers (stores) it in the management device 100 .

The AI registration unit 121 of this embodiment includes a selection tag generation unit 122 . The selection tag generator 122 generates, registers, and manages the AI selection tag information 154 . Details of the AI selection tag information 154 will be described later. Note that the selection tag generation unit 122 may be configured as a processing unit separate from the AI registration unit 121 .

The preset information management unit 123 generates, registers, and manages hyperparameters of registered AI models. Hyperparameters are algorithm tuning parameters applied to the AI engine, and include a plurality of parameter items and setting values for each item. For the hyperparameters, the AI vendor provides parameter items and setting values for each item, or the operator of the management device 100 generates setting values for each item based on the parameter items provided by the AI vendor. be able to. Then, in the present embodiment, a plurality of different hyperparameter information are generated in advance for the registered pre-learning AI model and stored in the hyperparameter preset information 152, as will be described later.

The learning processing unit 124 applies preset hyperparameters and causes the pre-learning AI model to perform learning processing using the learning data 162 . The inference processing unit 125 applies the inference data 163 and causes the learned AI model to perform inference processing. The learning processing unit 124 and the inference processing unit 125 can be provided with an analysis function that calculates the accuracy during learning (accuracy during learning) in the learning execution process and the accuracy during inference (accuracy during classification) in the inference execution process. . The calculated accuracy information is stored in the storage device 160 as learning accuracy information 164 and inference accuracy information 165, respectively. Any (well-known) method can be applied to the learning process, the inference process, the calculation process of the accuracy at the time of learning, and the calculation process of the accuracy at the time of inference.

Controller 130 primarily provides user management functions. The user registration unit 131 provides a predetermined user registration screen to the user terminal 300, and receives and registers user information of the AI-using user. The user information 161 includes the company name, department name, industry, company size, contact information, etc. to which the user belongs. Also, the user registration unit 131 issues user authentication information such as a user ID and password and stores it in the user information 161 . The user registration unit 131 can have a user authentication function using a user ID and password on a predetermined login screen.

Based on the screen information 171, the screen control unit 132 provides screens displayed on the user terminal 300, the vendor terminal 500, and the support provider terminal 600, and controls each screen. In addition, it cooperates with each processing unit and provides each terminal 300, 500, 600 with a screen corresponding to the function of each processing unit.

The upload data management unit 133 receives learning data 162 and inference data 163 prepared by AI users, and stores them in the storage device 160 for each user.

The AI model proposal unit 134 extracts one or more AI models using the AI selection tag information 154 associated with the registered AI models. Also, the AI model proposal unit 134 can have a recommendation control unit 135 . The recommendation function mainly uses various information such as the frequency of use of each AI model by all AI users, collaborative filtering between users, content-based filtering, etc., and selects the top AI model from among multiple AI models. Perform processing to extract AI models with high rates, AI models with high relevance, etc.

The AI model providing unit 136 performs a process of providing a user with a trained AI model based on learning data prepared by the AI user or a trained AI model that has undergone learning processing on the AI vendor side. Through the network, the user terminal 300 or AI-equipped equipment on the AI user side can be downloaded, or the learned AI model to be provided can be recorded in an electronic medium such as an HDD or flash memory and physically sent. can. In the latter case, the AI model providing unit 136 can have an output function to store the learned AI model selected by the AI user in a predetermined electronic medium. It should be noted that the function of downloading through a network and outputting to an electronic medium can be configured to have an export function of outputting in a format that can be executed by an AI-equipped device on the side of an AI user.

The user-specific AI management unit 137 provides various management functions through a user-dedicated management screen such as My Page. Details will be described later, but for example, on each management screen shown in FIGS. 19 to 25, AI models created and/or used on this platform, uploaded learning data, inference data, It is possible to manage the purchase of services, etc. Information and various histories (log information such as download history, purchase history, service usage history, etc.) input through these management screens are accumulated in the storage device 160 as AI management information 166 for each user.

The control device 140 mainly provides a function of introducing services provided by the support provider to the AI-using user and acting as an intermediary between the support provider and the AI-using user. As an example, the external service cooperation unit 141 provides a function of mediating annotation services. Annotation is a process of adding teacher data to learning data, and includes segmentation as necessary. For example, the external service providing unit 141 can provide an estimation function, an ordering function, etc., and the external service linking information 172 is, for example, fee information presented by the support company, an estimate, an order, a receipt, etc. Various information can be held.

The commercial transaction intermediation control unit 142 provides a function of purchasing AI models and services provided by AI vendors and services provided by support providers like an EC site. In other words, the AI-using user is provided with a purchase page displaying each product/service, and a commercial transaction function including payment is provided based on the AI-using user's purchase operation. The commercial transaction brokerage information 173 is information including various types of information such as information on products/services, charge information, and quotations. The commercial transaction brokerage control unit 142 provides a commercial transaction platform that allows AI users to directly purchase products and services from AI vendors and support providers. The services may include seminars and lectures such as AI-related education.

In addition, the external service cooperation unit 141 and commercial transaction brokerage control unit 142 can manage overall commercial transactions that occur on the AI platform of this embodiment. For example, usage fees for AI users who use this platform, AI vendors and support providers, fees incurred when AI users actually create AI models, usage fees for trained AI models, trained AI The controller 140 may also provide a payment function, such as for model purchase fees. The free/paid setting on the Platform is optional.

<Explanation of screen and each function>
When the user registration of the AI-using user is completed, the control device 130 generates a management screen (my page) for each user, and the AI-using user can "create ”/“Try it” AI development platform can be used.

FIG. 3 is a diagram showing an example of an AI model proposal screen provided to the user terminal 300. As shown in FIG. The AI model proposal screen includes two user interfaces (hereinafter referred to as UIs) of "Let's create an AI model (learning)" and "Let's use an AI model (inference)".

The UI "Let's create an AI model (learning)" applies learning data to the pre-learning AI model proposed by this system or selected by the AI user, and executes the learning process to generate the learned AI. Models can be created. Learning data is prepared by AI users and uploaded to this system.

The UI "Let's use an AI model (inference)" applies inference data to the trained AI model proposed by this system or selected by the AI user, executes inference processing, and outputs the inference result. Obtainable. Inference data is prepared by the AI user and uploaded to this system.

Then, the AI model proposing unit 134 of the present embodiment provides a proposal UI "Start from industry/role setting" using the AI selection tag information 154 and a proposal UI "Start from desired function selection".

FIG. 4 is an example of the AI selection tag information 154. As shown in FIG. The AI selection tag information 154 is information associated with each AI model registered by an AI vendor, and includes an AI model name, AI algorithm attribute information, industry, sub-industry, and application category (role). The linking of the AI selection tag information 154 to the AI model is performed by inputting and selecting each item of the AI selection tag information 154 when the AI vendor registers the AI model, or by the operator of this system. This can be done by inputting and selecting each item of the AI selection tag information 154 for the AI model.

FIG. 5 is a diagram showing an example of industries, and 10 industries are set as an example. Each industry has an industry sub-classification. In the example of FIG. 5, it is stipulated that the industry sub-classification "electrical machinery, automobiles, food, textiles, cosmetics, pharmaceuticals" is classified into the industry "manufacturing industry". The AI-using user can select an “industry” by referring to the industry sub-classification.

FIG. 6 is a diagram showing an example of usage categories. Usage categories are information corresponding to "roles", such as management/management, AI supervision/advanced, service/product development, research, production/quality control, sales/marketing, CS, logistics management, security, information system, and personnel. / Including general affairs, etc. The usage category is information for selecting what kind of usage (role) the AI user wants the AI model to have. You can make a selection based on your business or project.

FIG. 7 is a diagram showing an example of AI algorithm attribute information. The AI algorithm attribute information includes usage scene, category, algorithm type, and model configuration. The usage scene is, for example, information indicating in what kind of usage scene the AI model is good at being used.・Information such as the number of customers is linked. At this time, "Object Detection" is registered as the algorithm type, and "Single" is registered as the model configuration.

In addition to "Single", the model configuration includes "MIX", which represents a mode in which multiple single AI models are linked together and an AI model to which multiple algorithms are applied. For example, in the example of FIG. 7, the AI model that can utilize face/fingerprint authentication (seal) as a usage scene includes two AI algorithms, "Object Detection" and "Image Classification", and the model configuration is "MIX". is. Also, the AI model of the usage scene "recommendation for call center" includes three AI algorithms, and the model configuration is "MIX".

The AI model registered in this system is associated with AI selection tag information 154, and the AI user selects each tag of industry type, role (use category), and usage scene, so that the AI model is naturally suitable for the AI user. It has a mechanism for automatically selecting an AI model.

Although one or more AI models are selected, in particular, in this system, when an AI user creates an AI model using learning data in "Let's create an AI model", this system It provides an environment in which multiple AI models registered on the platform compete with each other and the most suitable AI model for the needs of AI users is presented.

The needs of AI users are diverse, and the evaluation criteria required for AI models are also different. In other words, if the problem to be solved and the purpose to be improved/improved by introducing an AI model are different, there are different evaluation criteria for each. For this reason, in the conventional mechanism of narrowing down the AI model to one or creating one by one and comparing them later, multiple AI models are " It was difficult to create an environment that presented the optimal AI model because it was not possible to "compete".

Therefore, in order to automatically present a plurality of AI models in the AI model selection function, AI selection tag information 154 is associated with each registered AI model, and a plurality of AI models are presented to the AI user. Preferably, the AI selection tag information 154 is constructed at the same time.

Then, while performing learning processing for a plurality of AI models, applying a plurality of different preset hyperparameters to each AI model, each trained AI model optimized for each AI model is By providing a "competitive mechanism" that compares and presents them collectively, we will realize an environment that presents the optimal AI model for the "direction" that AI users want for AI models.

FIG. 8 is a diagram showing a screen example of the proposal UI "start from setting of industry/role". "Your Industry" in Section 1 corresponds to the industries (including industry sub-categories) shown in FIG. "Your industry" in Section 2 corresponds to the application category (role) shown in FIG. "Usage scene" in Section 3 corresponds to the usage scene of the AI algorithm attribute information shown in FIG. Section 4 is an area where AI models selected through Sections 1 to 3 are displayed.

First, the AI model proposal unit 134 displays Section 1 “industry” in a selectable manner. The AI-using user selects at least one type of business and presses the confirm button. Triggered by pressing the confirmation button of Section 1, the AI model proposing unit 134 activates the display area of Section 2 and displays the "usage category" in a selectable manner. The AI-using user selects at least one application category and presses the confirm button. Triggered by the pressing of the confirmation button of Section 2, the AI model proposing unit 134 activates the display area of Section 3 and displays the "use scene" in a selectable manner. The AI-using user selects at least one usage scene and presses the confirm button. Triggered by the pressing of the confirmation button in Section 3, the AI model proposal unit 134 selects an AI using at least one type of industry, application category, and usage scene selected in each of Sections 1, 2, and 3 as selection conditions. Search tag information 154 . Then, an AI model that satisfies the three selection conditions is extracted, the display area of Section 4 is activated, and the extracted AI model is displayed.

It should be noted that the control of each display area of Sections 2 and 3 may be configured such that, for example, the content selected in the previous Section is used as a selection condition to narrow down the display content of the next Section. For example, the AI model proposal unit 134 searches the AI selection tag information 154 for the "use category" displayed in Section 2 using the "industry" selected in Section 1 as a search condition, triggered by pressing the confirmation button in Section 1. , can be controlled to extract and display only the application category linked to the corresponding AI model. The same is true for Section 3. Triggered by the pressing of the Confirm button in Section 2, an AI model that is in the industry extracted in Section 1 and is linked to the application category selected in Section 2 is searched from the AI selection tag information 154, and the corresponding It is also possible to control so that only usage scenes linked to the AI model to be used are extracted and displayed in Section 3 so that they can be selected.

In this embodiment, the AI selection tag information 154 is searched using three selection conditions (industry, application category, usage scene) selected by the AI user, and a plurality of corresponding AI models are presented as the search results.

FIG. 9 is a diagram showing a registration screen for learning data. An AI user uploads learning data to this system when creating an AI model. The upload data management unit 133 accepts learning data through the screen shown in FIG. 9 and stores it in the storage device 160 . At this time, the AI user can make an annotation request for the uploaded learning data. For example, when an annotation request button is pressed, a list of support companies that perform annotations and information about fees are displayed on the user terminal 300, and the AI-using user can request annotations from the support companies. In this case, as described above, it may be configured to have a function for performing exchanges between the AI-using user and the support business according to each stage of commercial transactions such as quotation, ordering, and settlement.

Also, since the learning data registration screen in FIG. 9 is intended to specify learning data for performing learning processing, it can be uploaded from the AI user (user terminal 300), and can also be uploaded within the system ( Learning data already stored in the storage device 160) can also be specified. For example, learning data can be uploaded in advance and registered on the Data management screen shown in FIG. The learning data includes learning data previously uploaded and stored by the AI user and learning data annotated through an annotation request.

FIG. 10 is a screen displaying an AI model being created. The learning processing unit 124 executes learning processing when learning data is specified in FIG. The execution status of the learning process can be confirmed in FIG. For example, it is possible to display the time required to create an AI model, or the remaining time until creation is completed. The learning processing execution status is provided from the learning processing unit 124 . It is also possible to stop model creation. When model creation is selected to be canceled, the learning processing unit 124 cancels the learning process.

FIG. 11 is an example of a screen displaying learning results of the created AI model. The learning processing unit 124 generates learning accuracy information in the learning process. As shown in FIG. 11, the AI model proposing unit 134 displays learning accuracy information of each learned AI model. The AI-using user can select a trained AI model, and display information (a graph of prediction results during learning, prediction accuracy, etc.) of the selected trained AI model. In addition, the total number of models verified by learning (the number of different hyperparameter presets) is also displayed.

The screen of FIG. 11 can be displayed by switching between "Summary" and "Detail" using tabs. FIG. 12 is a diagram showing a mode in which "Detail" is displayed, in which detailed information (learning result table items and their results) of the selected trained AI model is displayed. Detailed information includes various evaluation items such as elapsed, last_epoch, last_val_acc, last_val_loss, last_acc, last_loss, batch_size, learning_rate, patience, and epochs.

Note that the learning accuracy and the learning result are arbitrary information according to the AI engine and algorithm, and known techniques can be applied. For AI models that perform the same image processing, unified evaluation items and learning results regarding image processing are provided, and even for AI models that perform the same image processing, different evaluation items from other AI models can be provided. As an example, for each category of function classification master information described later, different evaluation items and learning results are provided between categories, and each AI model belonging to the category can provide evaluation items and learning results unified by category. can.

FIG. 13 is a diagram showing an example of a screen of a proposal UI "Start by selecting a function to use" in the UI "Let's create an AI model (learning)". Algorithms in AI models can be classified by function. For example, the AI algorithm attribute information in FIG. 7 can be configured to include "categories" for algorithm types, and the "categories" represent function classifications. For example, function classification master information can be set in advance and controlled to be associated with each AI model in the AI selection tag information 154 .

Functional classification master information, for example, CV (Computer Vision), NPL (Natural Language Processing), ASR (Automatic Speech Recognition), including category divisions such as time series processing In addition, each AI model belonging to the functional classification "CV" has algorithms for object detection, image classification, posture estimation, etc. based on image recognition, so for example, an AI model belonging to the functional classification "CV" can be further categorized by each algorithm type.In the example of FIG. The same applies to other function classifications.

Based on the function classification master information and the AI selection tag information 154, the AI model proposal unit 134 provides the function-specific selection screen shown in FIG. 13, and selects a plurality of AI models belonging to the selected function.

Then, when a function is selected in FIG. 13, similarly to the proposal UI "start from setting of industry/role", the screen for specifying learning data in FIG. 9, the screen for displaying the AI model being created in FIG. Each screen is provided in order of the learning result display screen of FIG. 11 and FIG. 12, and each corresponding process is performed.

Next, the UI "Let's use an AI model (inference)" will be described. The AI model to be used for inference can also be selected from the UI "Let's use an AI model (inference)" and the proposal UI "Start by setting the industry/role" or the proposal UI "Start by selecting the functions you want to use". However, the AI model selected at this time can be controlled to be one.

In this case, for multiple AI models selected from the interface of the proposal UI "start from industry/role setting" or the proposal UI "start from selection of functions you want to use", the AI user can select any of the multiple AI models. can be controlled so that one AI model can be selected by the AI model proposal unit 134, or the AI model proposal unit 134 can be controlled to automatically select one AI model.

In the selection and automatic selection by AI users, for example, ranking based on the usage frequency of AI users, collaborative filtering, content-based filtering, etc. can be applied, and usage frequency, selection history of other AI users, etc. can be presented as reference information, or an AI model with the highest usage frequency or an AI model selected by another AI user with the highest preference similarity can be automatically selected.

In addition, the UI "Let's use an AI model (inference)" is also configured to apply one inference data to multiple trained AI models and obtain multiple inference results for each trained AI model. You can also

FIG. 14 is a diagram showing an inference data specification screen in the UI "Let's use an AI model (inference)". This is the same as the learning data registration screen shown in FIG. 9, and the AI user uploads inference data to this system in order to make an inference using the AI model. It is also possible to designate a plurality of different data for inference and perform inference processing on the trained AI model.

FIG. 15 is a screen that displays the state during inference. When the inference data is designated, the inference processing unit 125 executes inference processing using the selected trained AI model. If multiple pieces of inference data are specified, inference processing is performed as many times as the number of pieces of inference data. The inference processing unit 125 outputs the execution status (remaining time) of the inference processing so that it can be confirmed on the screen as shown in the example of FIG.

FIG. 16 is a screen for displaying inference results, and accuracy information (certainty, etc.) is displayed for each designated inference data. The example of FIG. 16 is a "Summary" screen, which displays a list of accuracy information for each designated inference data. FIG. 17 shows a "Detail" screen that can be displayed by switching tabs on the screen that displays the inference results, and a detailed inference result for one piece of inference data is displayed on the "Detail" screen. The "Detail" screen, for example, displays each object recognized in the image data as the analysis result, displays the image of the inference data, and frames each recognized object in the image (using the classification color of the analysis result). It can be displayed by enclosing it with a frame of a different color for each object used.

FIG. 18 is a screen displaying the inference result of an AI model whose algorithm type is or includes time series. Prediction results are displayed in the order of inference data entered in chronological order. It should be noted that although the mode in which prediction result 1 and prediction result 2 are displayed is shown, this is arbitrary according to the characteristics of the algorithm of the AI model.

19 to 24 are diagrams showing user management screens provided to AI users who use this system.

In the left column of the screen in FIG. 19, menus for project management, data management, AI management, and optional services are displayed. Project management is a management screen for organizing what kind of AI model is used for what purpose. can be tied.

FIG. 20 is a list screen of AI models used in each project. AI model name, algorithm, data name (learning data name), related project name, data capacity of AI model, etc. are displayed.

FIG. 21 is a diagram showing an example of a data management screen. The Data management screen provides functions for uploading, deleting, and downloading learning data and inference data. In addition, the learning data specified from the above-mentioned "Let's create an AI model" UI and the inference data specified from the "Let's use an AI model" UI can be managed on this Data management screen.

In the example of FIG. 21, along with the data name and description (outline) of the data, as details of the data, whether the data is for inference or learning, and what kind of AI model the registered learning data and inference data are. You can select and enter whether it is applied to (category name or algorithm type). In addition, it is also possible to manage as a log whether or not the learning data (or inference data) has been delivered from the support company after an annotation request.

The data management screen can be used as a data management environment for the purpose of requesting annotations, for example, regardless of whether the data is used in AI model learning processing and inference processing on this system. can. Training data and inference data for AI models that operate outside this system (irrespective of AI models created or purchased by this system, or AI models prepared by AI users independent of this system) You can download it and use it.

In addition, in this system, a predetermined amount of storage area is allocated to an AI user, so that learning data, inference data, and an AI model can be stored on the management device 100 side. In the example of FIG. 21, the storage usage rate is displayed, and management is performed so as not to exceed the storage capacity by deleting or downloading as necessary. Note that the capacity of each storage area for Data management and AI model management can be set individually or collectively in total.

FIG. 22 is an AI management screen. In the example of FIG. 22, an AI model created by this system and an AI model used are displayed. Various information such as AI model name, algorithm, creation date, created user name, data name, related project, etc. can be input. Also, for each AI model, it is possible to check whether it is a model for inference or a model for learning, and the status is also displayed. The status displays states such as learning (unavailable) and model available.

FIG. 23 shows an example of a purchase screen when downloading and purchasing an AI model on the AI model management screen shown in FIG. When the download button in FIG. 22 is pressed, the "valid download remaining number" in FIG. 23 is displayed. The remaining valid number is the upper limit of the number of downloads that can be made. It can be purchased (downloaded) by selecting "Additional purchase". It is also possible to specify and purchase one or more AI models.

FIG. 24 shows a model being created in the AI model management shown in FIG. You can manage the learning status of the AI model you are creating on the UI of "Let's create an AI model", and as shown in FIG. Time (including time remaining bar), etc.

For AI models that have already been trained, the learning results shown in FIG. 11 and/or FIG. 12 can be displayed by pressing the "Check learning results" button. Furthermore, the AI model being created (learning) is controlled so that a "stop model creation" button can be selected, and model creation can be stopped from the AI model management screen.

FIG. 25 is an optional service management screen. The example of FIG. 25 is illustrated in the form of a product/service purchase page on an EC site. As described above, AI vendors and support providers can have AI users purchase products and services through this platform. AI package models, annotation services, education services, etc. provided by AI vendors and support companies are displayed, and the most popular! It is also possible to display recommended products such as Purchase histories can also be managed for each user, and can be controlled to be viewable on the optional service management screen.

<Description of the function that presents the optimal AI model>
This embodiment has a function of presenting an optimal AI model. This function consists of two elements. The first is to extract an AI model suitable for the AI user from the industry, role, and usage scene using the AI selection tag information 154 described above. As a result, even if the AI user does not have specialized knowledge, an AI model that matches him/herself is automatically presented.

The second is a mechanism that allows multiple AI models to "compete" with respect to the "direction" that AI users want from AI models. FIG. 26 is a diagram for explaining a mechanism for "competing" a plurality of AI models. As shown in FIG. 26 , a plurality of different hyperparameter groups are prepared in advance for the registered pre-learning AI model and held as hyperparameter preset information 152 . For example, in machine learning and deep learning, hyperparameters such as batch size, number of iterations, and number of epochs are known.

The batch size is the number of data included in each subset when the data set is divided into a plurality of subsets. The number of iterations is the number of times of learning necessary for each data in the data set to be used for learning at least once, and is automatically determined by determining the batch size. If the batch size is 10 with 80 data sets, the number of iterations will be 8, and the learning will be repeated 8 times. In addition, the number of epochs defined how many times to divide the data set into subsets according to the batch size, set each data in the data set to learning at least once as one cycle, and how many times to execute this one cycle.

In this embodiment, as a plurality of different hyperparameter groups, hyperparameters 1 to N with different batch sizes, iteration numbers, epoch numbers, etc. are prepared in advance, and hyperparameter 1 is applied to one pre-learning AI model 1. Learning model 1-1 (preset 1-1) learned by applying hyperparameter 2 (preset 1-2), learning model 1-2 (preset 1-2) learned by applying hyperparameter N A trained learning model 1-N (preset 1-N) is generated.

This is applied to each of a plurality of AI models selected according to the type of industry, role, and usage scene using the AI selection tag information 154, and a plurality of different trained AI models are generated for one AI model.

Here, the major difference from the conventional mechanism will be described. Conventionally, as described in Patent Document 2, the optimum hyperparameter is “searched”, and the hyperparameter is gradually changed to search for the learning parameter with the highest performance. However, such a conventional method is an optimization of one AI model, and even if automated, it takes time and cost.

On the other hand, this system provides a mechanism to make multiple AI models “compete” in response to the “direction” that AI users seek for AI models, and each of the multiple different preset hyperparameters is nullified. As shown in FIG. 26, each of a plurality of different AI models that are applied to conditions is subjected to learning processing for the number of different preset hyperparameters. That is, for one AI model, a plurality of trained AI models with different learning parameters are generated at once, and a trained AI model with good performance is extracted from among them. Then, in a plurality of AI models, the training results of the trained AI models with good performance are collectively provided.

In this embodiment, instead of optimizing one AI model to meet the needs of a wide variety of AI users, multiple AI models that broadly cover the evaluation criteria required for AI models are provided in a “competitive” environment. By doing so, it is possible to prepare many options for the “direction” that the AI user wants for the AI model. As a result, it is possible to reduce the time and cost required to search for hyperparameters, etc., and to provide an environment in which multiple AI models compete with each other in response to various needs of AI users.

Therefore, even if an AI user does not have specialized knowledge, an AI model that matches him/herself is automatically presented, and it is easy to propose an AI model that meets various needs by having multiple AI models compete. Therefore, it is possible to realize an environment in which users without specialized knowledge can create and test AI models that meet their needs.

<Description of the linkage interface function>
FIG. 27 is a diagram for explaining the cooperation interface function. The link interface function is a support function for constructing a linked AI model in which different AI models are linked. It can be applied to aspects of the composite model described above.

For example, as shown in FIG. 30, when an AI model for object detection and an AI model for object identification are separately provided (registered), each AI model has different output data and input data formats. So you can't simply concatenate. Therefore, the control device 120 includes a linkage interface control unit 126 and provides an interface module that converts the output data of the AI model on the upstream side of the connection into the input data of the AI model on the downstream side of the connection. As a result, it is possible to create a connected AI model in any combination.

As linkage information, the storage device 150 holds information (interface information) on each data format of input data and output data of an AI model provided by an AI vendor as model linkage information 155 .

The cooperation interface control unit 126 refers to the model cooperation information 155 and provides an interface module between each AI model according to the connection order from upstream to downstream between two or more AI models to be connected. The interface module can be generated in advance and stored in the model linkage information 155, or can be controlled to be generated by the linkage interface control unit 126 each time a linked AI model is generated.

<Description of the processing flow of this system>
28 to 30 are diagrams showing the processing flow of this system. As shown in FIG. 28, the AI vendor performs vendor registration through the vendor terminal 500 (S501). The management device 100 performs vendor registration acceptance processing and registration processing (S101). After registration, the AI vendor updates the pre-learning AI model and/or the learned AI model to the management device 100 and registers them (S502).

When the AI vendor registers the AI model, the management device 100 stores the program module of the AI model in the storage device 150 (S102). AI models registered by AI vendors are trained AI models and pre-learned AI models.

The management device 100 generates the hyperparameter preset information 152 of the stored (registered) AI model, and stores it in association with the AI model (S103). At this time, as described above, a plurality of different hyperparameter groups are set for one AI model. In other words, the hyperparameter preset information 152 is stored for repeating the learning process by the number of set hyperparameters different from each other and generating a plurality of trained AI models corresponding to the number of hyperparameters. As described above, each hyperparameter that constitutes the hyperparameter group can be generated and provided by the operator of this system or the AI vendor.

On the other hand, the management device 100 performs generation and registration processing of the AI selection tag information 154 for the registered AI model (S104). All AI models registered in this system are categorized by AI selection tag information 154 by AI algorithm attribute information, industry, sub-industry, application category (role), etc. shown in FIG.

Next, the AI-using user performs user registration for this system (S301, S105). The AI-using user signs in to the system from the user terminal 300, and the management device 100 performs user authentication (S302, S106).

After user authentication, the management device 100 causes the user terminal 300 to display, for example, a user-dedicated page shown in FIG. 3 (S107). First, the UI "Let's create an AI model (learning)" for learning processing will be described. As shown in FIG. 28, when the proposal UI “Start from industry/role setting” is selected in the UI “Let’s create an AI model (learning)” (S303, S304, S305), the management device 100 AI proposal processing through the screen of No. 8 is performed (S108).

The management device 100 refers to the AI selection tag information 154 and selects a plurality of corresponding AI models (learning target AI models) based on the business type, application category, and usage scene selected by the AI user (S305, S108, S109). The management device 100 receives learning data uploaded by the AI user (S306, see FIG. 9).

AI-using users can use annotation services provided by support providers for uploaded learning data. For example, in the screen example of FIG. 9, the request can be made by pressing the "Request Annotation" button. The management device 100 determines whether or not there is an annotation request in the learning data reception process (S110). If there is no annotation request, the process proceeds to step S113 and stores the received learning data for each user.

When it is determined in step S110 that there is an annotation request, the management device 100 performs annotation request processing (S111). The annotation request process is a process of requesting an estimate from the support company, presenting the estimate received from the support company to the AI user, and mediating the ordering by the AI user (S307, S111, S601). Upon receiving an order from the AI user, the management device 100 provides the uploaded learning data to the support company, and the support company performs annotation (S112, S602). The support provider transmits the annotated learning data (for example, learning data with a teacher label attached) to the management device 100 . The management device 100 stores the annotated learning data in the storage device 160 (S113). The annotation request can be made at any timing on the management screen after uploading the learning data (see FIG. 21, etc.).

FIG. 29 is a processing flow following FIG. In the UI "Let's create an AI model (learning)", multiple AI models (pre-learning AI models) are selected based on the AI selection tag information 154, and learning data (including annotated learning data) is prepared. Then, the management device 100 executes learning processing.

The management device 100 executes learning processing for each of the selected AI models (S151). The management device 100 sets one hyperparameter from a group of preset hyperparameters associated with the AI model to be learned (S152), and performs learning processing using learning data (S153). The management device 100 generates learning accuracy information associated with the learning process (S154), and stores the learned AI model and the learning accuracy information (S155).

The management device 100 confirms whether or not learning processing has been performed with all hyperparameters (S156), sets the next hyperparameter, and repeats steps S152 to S155 by the number of preset hyperparameters. do. Then, the processing from step S152 to step S157 is executed for all of the selected AI models. In other words, after the learning process for one AI model is completed, it is checked in step S158 whether or not there is any remaining learning process for the other selected AI model, and the next AI model is set as the learning target ( S159), the processing from S152 to step S157 is repeatedly executed.

The management device 100 extracts an AI model with the highest accuracy during learning from among a plurality of trained AI models for different hyperparameters generated for one AI model. This is performed for each of the selected AI models (learning target AI models), and extracted as an optimal AI model group (S160, FIG. 26).

Then, the management device 100 provides the user terminal 300 with learning results (including accuracy information during learning) including the AI model group with the highest accuracy during learning extracted from each of the selected AI models (S161 , FIG. 11, etc.). At this time, as shown in FIG. 11, it is possible to perform ranking in descending order of accuracy during learning, or display the accuracy during learning as a graph (FIGS. 11 and 12).

The AI-using user can select a trained AI model included in the learning result and execute inference processing. The AI user uploads inference data to be applied to the selected trained AI model (or designates pre-uploaded inference data), and transmits an inference process execution request to the management device 100 (S310). , S311).

When receiving the inference processing execution request, the management device 100 executes inference processing by applying the specified inference data to the selected trained AI model, and generates inference accuracy information (S162). The generated inference accuracy information is provided to the AI user as an inference result (S163).

For example, the management device 100 receives a download request for an AI model by selecting a download button in FIG. 11 or the like or in response to a download operation on the management screen shown in FIG. model) is provided (S312, S164).

FIG. 30 is a diagram showing the flow of inference processing in the UI "Let's use an AI model (inference)", and shows branch processing from step S303 in FIG. As shown in FIG. 30 , when the proposal UI “start from selection of desired functions” is selected in the UI “Let’s use AI model (inference)” (S303a, S3051), the management device 100 displays the screen in FIG. AI proposal processing through is performed (S1081).

The management device 100 refers to the AI selection tag information 154 based on the function selected by the AI user, and selects a plurality of corresponding learned AI models (S3051, S1081). The selected one or more learned AI models are provided to the user terminal 300 as AI model proposal results (S1091). At this time, if there are multiple trained AI models to be proposed, the AI-using user can select one or more trained AI models from the provided AI model proposal results (S3052). The management device 100 receives selection information from the AI user and determines a learned AI model for performing inference processing (1092).

Next, the management device 100 receives inference data uploaded by the AI user (S3110, see FIG. 14).

The management device 100 performs inference processing by the selected trained AI model using the received inference data (see FIG. 15), and generates inference accuracy information 165 (S1620). As a result of the inference processing, the management device 100 provides the inference-time accuracy information 165 to the user terminal 300 (S1630, see FIGS. 16 to 18).

For example, on the management screen shown in FIG. 23, the management device 100 provides the corresponding trained AI model (inference AI model) upon receiving a download request for an AI model (S3120, S1640).

In the example of FIG. 30, the proposal UI "start from selection of functions you want to use" was explained as an example. can be applied. Conversely, in the example of FIG. 28, it is also possible to execute the UI ``Let's create an AI model (learning)'' by the proposed UI ``Start from selection of functions you want to use''.

<About the AI platform realized by this system>
The AI platform system of this embodiment is a system that specializes in the convenience of AI users (everyone can use it). It is possible to realize an environment where you can create and try

(A) Based on the AI selection tag information 154, the best AI model (pre-learning AI model and learned AI model) can be automatically selected from the AI user's industry, occupation, and the like. As a result, even if the user is not an engineer or does not have much knowledge about AI, he or she can select an AI model that meets the needs of the site or business of the user using AI.

(B) Multiple AI models can be registered for one AI category (one function classification such as image processing and object detection) to handle a wide variety of data, and managed by AI selection tag information 154. do. In other words, a plurality of AI models proposed to the AI-using user are controlled to be extracted based on the AI selection tag information 154, and an environment is formed in which a plurality of AI models are competitively provided.

(C) Then, as a real environment in which a plurality of AI models compete, different hyperparameter groups are preset in advance, and in the learning process of one AI model, a plurality of trained AI models corresponding to the number of hyperparameters are generated. do. As a result, it is possible to reduce the time and cost required to search for hyperparameters (learning parameters), etc., and create an environment where AI models can be created easily, rather than seeking optimization of AI models.

(D) The platform can support the entire process of creating an AI model, testing it, and putting it into practical use. For example, AI education introduction → data preprocessing (annotation) support → data registration → AI model selection (choose from job type and function) → selection of optimal AI model from learning results → download of trained AI model (AI model for inference) (including application to the implementation environment on the user side) can be centrally managed and operated on this platform. In addition, this platform can also be configured to have functions such as introducing vendors who build implementation environments on the user side, and can provide overall support including utilization in actual sites and businesses.

(E) In particular, in (D) above, conventionally, when introducing AI, we understand AI, prepare data, and develop systems according to the usage scene from the launch of the AI utilization concept. We used to pay a large amount of cost including time until AI introduction, such as selecting a company, but by using this platform, there is no need to spend time and money, and user convenience for AI introduction will be improved. can be done.

The embodiments of the present invention have been described above. It can also be constructed by distributing the parts arbitrarily,

Also, the user terminal 300, vendor terminal 500, and support provider terminal 600 are desktop computers or tablet computers, and have data communication functions and calculations through IP (Internet protocol) networks or mobile communication networks. It has functions (CPU, etc.). In addition, a display device (or a touch panel type display device), an input means such as a keyboard, and an auxiliary storage device such as an HDD or SSD (solid state drive) can be provided. The mode of connection between the management device 100 and each terminal 300, 500, 600 may be either wireless communication or wired communication. 100, it is desirable to establish a stable communication environment.

Each function of the management device 100 can be realized by a program, and a computer program prepared in advance for realizing each function is stored in the auxiliary storage device, and a control unit such as a CPU is stored in the auxiliary storage device. The program read out to the main storage device is read out, and the program read out to the main storage device is executed by the control unit, so that the function of each unit can be operated.

The program can also be provided to the computer in a state recorded on a computer-readable recording medium. Examples of computer-readable recording media include optical discs such as CD-ROMs, phase-change optical discs such as DVD-ROMs, magneto-optical discs such as MO (Magnet Optical) and MD (Mini Disk), floppy (registered trademark) discs, and the like. Examples include magnetic disks such as removable hard disks, compact flash (registered trademark), smart media, SD memory cards, memory cards such as memory sticks. Further, a hardware device such as an integrated circuit (IC chip, etc.) specially designed and configured for the purpose of the present invention is also included as a recording medium.

100 AI platform device (management device)
110 Communication device 120 Control device 121 AI registration unit 122 Selection tag generation unit 123 Preset information management unit 124 Learning processing unit 125 Inference processing unit 126 Cooperation interface control unit 130 Control device 131 User registration unit 132 Screen control unit 133 Upload management unit 134 AI Proposing unit 135 Recommendation control unit 136 AI model providing unit 137 User-specific AI management unit 140 Control device 141 External service cooperation unit 142 Commercial transaction brokerage control unit 150 Storage device 151 Registered AI model DB
152 Hyperparameter preset information 153 Trained AI model 154 AI selection tag information 155 Model linkage information 160 Storage device 161 User information 162 Data for learning 163 Data for inference 164 Accuracy information during learning 165 Accuracy information during inference 166 AI management information for each user 170 Storage device 171 Screen information 172 External service cooperation information 173 Commercial transaction brokerage information 300 User terminal 500 Vendor terminal 600 Support provider terminal

Claims (6)

An AI platform system that provides a platform for AI users to create and test AI models that meet their needs,
a first storage unit for storing AI selection tag information associated with each AI model provided by an AI vendor, including business type, application, and usage scene;
A second storage unit that stores hyperparameter preset information, which is AI model learning preset information provided by an AI vendor and is generated for each AI model and configured by a hyperparameter group including a plurality of mutually different hyperparameters. When,
The AI user controls the industry, application and usage scene to be selectable through the user terminal, searches the AI selection tag information using the selected industry, application and usage scene, and selects the corresponding AI model as the learning target AI model. an AI model proposal unit that extracts a plurality of as
a learning processing unit that generates a trained AI model by executing learning processing of the learning target AI model using learning data specified by an AI user;
an inference processing unit that uses inference data specified by an AI user to execute inference processing of a trained AI model to generate an inference result;
The learning processing unit
First learning for generating a plurality of trained AI models for each hyperparameter by setting each hyperparameter of the hyperparameter group linked to the learning target AI model as a learning parameter, and performing learning processing for the number of the hyperparameters. while performing processing, executing the first learning processing for all of the plurality of extracted learning target AI models,
The AI model proposing unit, for each of the plurality of learning target AI models, has learned from a plurality of hyperparameter-specific trained AI models generated based on the learning accuracy generated along with the first learning process. An AI platform system characterized by extracting an AI model and providing an AI user with an accuracy at the time of learning of each of the extracted trained AI models as a learning result of each of the plurality of learning target AI models. The AI model proposing unit causes the user terminal to display a selection screen including selection areas for industry, application, and usage scene based on the AI selection tag information, and displays the selection screen for the industry, application, and usage scene through the selection screen. providing the user terminal with an upload screen for accepting learning data to be applied to the plurality of learning target AI models extracted based on each selection information of from the AI user,
2. The learning processing unit according to claim 1, wherein the learning data specified through the upload screen is used to perform learning processing on the plurality of learning target AI models extracted through the selection screen. AI platform system. 2. The AI model proposing unit generates a learning result screen in which the learning results of each of the plurality of AI models to be learned are ranked in descending order of accuracy at the time of learning, and provides the learning result screen to the user terminal. Or the AI platform system according to 2. Functional classification information based on AI algorithm characteristics is associated with each AI model provided by the AI vendor,
The AI model proposing unit provides the user terminal with a function selection screen based on the function classification information, and extracts an AI model linked to the function classification selected on the function selection screen as the learning target AI model. The AI platform system according to any one of claims 1 to 3, characterized by: Further comprising an external service cooperation unit that accepts annotation requests for learning data from AI users and provides the annotation requests to support providers that implement annotations,
The external service cooperation unit provides the learning data received from the AI user to the support business operator, and receives the annotated learning data from the support business operator and assigns learning data to each user. 5. The AI platform system according to any one of claims 1 to 4, wherein the AI platform system is stored in a storage area. A program executed by a computer device that provides a platform for AI users to create and try out AI models that meet their needs, wherein the computer device comprises:
a first function of storing AI selection tag information associated with each AI model provided by an AI vendor, including the type of business, application, and usage scene;
a second function of storing hyperparameter preset information, which is AI model learning preset information provided by an AI vendor and which is generated for each AI model and configured by a hyperparameter group including a plurality of mutually different hyperparameters; ,
The AI user controls the industry, application and usage scene to be selectable through the user terminal, searches the AI selection tag information using the selected industry, application and usage scene, and selects the corresponding AI model as the learning target AI model. A third function that extracts a plurality of as
a fourth function for generating a trained AI model by executing a learning process for the learning target AI model using learning data specified by an AI user;
a fifth function of generating an inference result by executing an inference process of the trained AI model using the inference data specified by the AI user;
The fourth function is
First learning for generating a plurality of trained AI models for each hyperparameter by setting each hyperparameter of the hyperparameter group linked to the learning target AI model as a learning parameter, and performing learning processing for the number of the hyperparameters. while performing processing, executing the first learning processing for all of the plurality of extracted learning target AI models,
The third function is a learned AI from among a plurality of hyperparameter-specific trained AI models generated based on learning accuracy generated along with the first learning process, for each of the plurality of learning target AI models. A program for extracting a model and providing an AI user with an accuracy at the time of learning of each of the extracted trained AI models as a learning result of each of the plurality of learning target AI models.

Images