Information processing device, control method and program for information processing device

Abstract

To provide a technique that reduces the risk of AI ethics issues arising from AI models made public by users.SOLUTION: An information processing device capable of communicating with another information processing device includes: acquisition means for acquiring information on a trained AI model from the other information processing device; evaluation means for evaluating the trained AI model; and determination means for determining a disclosure range to which the information processing device will disclose the trained AI model based on an evaluation result of the evaluation means.SELECTED DRAWING: Figure 4

Description

[Technical Field]

[0001] The present invention relates to an information processing device, a control method for an information processing device, and a program, and more particularly to a technology for learning a plurality of AI (Artificial Intelligence) tasks. [Background technology]

[0002] As AI becomes more widespread, there are concerns that AI's judgment results may result in discrimination based on race, gender, etc., or that AI may mistakenly identify humans as animals, and as a countermeasure, efforts are underway worldwide to formulate AI ethics. Furthermore, AI ethics may differ depending on the country or region.

[0003] In recent years, AI models are increasingly being created not only by companies and organizations but also by ordinary users for their own purposes. In addition, there are internet-based services that allow ordinary users to publish and obtain machine-learned AI models they have created, making it possible for many users to use a published model. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent Publication No. 2021-33707 Summary of the Invention [Problem to be solved by the invention]

[0005] However, AI models created by users may not fully comply with AI ethics. In such cases, if the AI ​​model is made public on a server and available to users around the world, the output of the AI ​​model may cause problems in the marketplace that violate AI ethics. For example, a model that autofocuses on a specific person may end up autofocusing on other animals.

[0006] The present invention has been made in consideration of the above-mentioned problems, and aims to provide technology that reduces the risk of AI ethics issues arising from AI models made public by users. [Means for solving the problem]

[0007] To achieve the above object, an information processing device according to the present invention comprises: An information processing device capable of communicating with other information processing devices, An acquisition means for acquiring information about a trained AI model from the other information processing device; A determination means for identifying evaluation items for AI ethics based on information about the trained AI model and determining evaluation data according to the evaluation items; a calculation means for calculating the evaluation accuracy of the trained AI model based on the evaluation data; Based on the evaluation accuracy of the trained AI model and a reference value associated with the evaluation item, A determination means for determining a disclosure range in which the information processing device will disclose the trained AI model; Equipped with When the evaluation accuracy is less than the reference value and a difference between the evaluation accuracy and the reference value is equal to or less than a preset limit value, the decision means decides to make the trained AI model public to evaluation users, which are a portion of users who evaluate the trained AI model; The information processing device includes: an acquisition means for acquiring a review result of the trained AI model by the evaluation user; a change means for changing the reference value based on the review result; Further It is characterized by being equipped with: [Effects of the Invention]

[0008] According to the present invention, it is possible to reduce the risk of AI ethics issues arising from AI models made public by users. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 2 is a hardware configuration diagram of a server-side information processing device in an information processing system according to an embodiment. [Figure 2] FIG. 2 is a hardware configuration diagram of a user-side information processing device in the information processing system according to an embodiment. [Figure 3] FIG. 1 is a configuration diagram of an information processing system according to an embodiment. [Figure 4] FIG. 1 is a functional block diagram of an information processing system according to an embodiment. [Figure 5] FIG. 4 is a functional block diagram of a disclosure range determination unit according to an embodiment. [Figure 6] 1 is a flowchart showing a procedure of a process executed by an information processing apparatus according to an embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. Note that the following embodiments do not limit the scope of the claimed invention. Although multiple features are described in the embodiments, not all of these multiple features are necessarily essential to the invention, and multiple features may be combined arbitrarily. Furthermore, in the accompanying drawings, the same reference numerals are used to designate the same or similar components, and redundant explanations will be omitted.

[0011] (First embodiment) In this embodiment, an example is described in which information about a trained AI model is obtained from a user-side information processing device and evaluated, and based on the evaluation results, the server-side information processing device determines the disclosure range to which the trained AI model will be made public.

[0012] <Server side hardware configuration> FIG. 1 is a hardware configuration diagram of an information processing device 100 on the server side in the information processing system according to this embodiment.

[0013] The information processing device 100 includes a processor 101, a storage medium 102, a memory 103, and an interface 104, and these components are connected via a system bus 105. The information processing device 100 is configured on a cloud server.

[0014] The processor 101 is, for example, one or more CPUs, and controls various devices connected to a system bus 105. The storage medium 102 is, for example, a ROM, and stores a BIOS (Basic Input / Output System) program and a boot program. The memory 103 is, for example, a RAM, and is used as the main storage device for the processor 101, which is a CPU. The interface 104 communicates information via a network. The communication interface may be Ethernet (registered trademark), or any type of interface may be used, such as USB, serial communication, or wireless communication.

[0015] <User hardware configuration> FIG. 2 is a hardware configuration diagram of an information processing device 200 on the user side in the information processing system according to this embodiment.

[0016] The information processing device 200 includes a processor 201, a storage medium 202, an input device 203, a display device 204, a memory 205, a hard disk 206, a media drive 207, and an interface 209, and each component is connected via a system bus 209. The information processing device 200 is a terminal device on which a user views and operates a screen, such as a PC (personal computer) or a tablet terminal.

[0017] The processor 201 is, for example, one or more CPUs, and controls various devices connected to a system bus 209. The storage medium 202 is, for example, a ROM, and stores a BIOS program and a boot program. The input device 203 is, for example, a touch panel, keyboard, mouse, or robot controller, and performs processing related to input of information, etc. 204 is a display device that outputs the results of calculations by the information processing device 200 in accordance with instructions from the CPU 201, and display information transmitted from the server-side information processing device 100. The display device 204 may be a liquid crystal display device, a projector, an LED indicator, or the like, and any type of device may be used.

[0018] The memory 205 is, for example, a RAM, and is used as the main storage device of the processor 201. The hard disk 206 is used to save and load application programs, data, libraries, etc. The media drive 207 makes it possible to write data from the hard disk 206 to a removable storage medium, thereby making it possible to transfer the written data to an external digital still camera, PC, or tablet device.

[0019] The interface 208 communicates information via a network, for example, with the server-side information processing device 100. The communication interface may be Ethernet, or any type of interface such as USB, serial communication, or wireless communication.

[0020] <Configuration of information processing system> Fig. 3 is a configuration diagram of an information processing system according to this embodiment. As shown in Fig. 3, a server-side information processing device 100 is connected to multiple user-side (client-side) information processing devices 200, and controls the transmission and reception of data and display information.

[0021] In this embodiment, an example is described in which a user determines whether an AI model trained by a user sufficiently complies with AI ethics (to the extent that the likelihood of future problems is low, not to guarantee 100%), and determines which users to publish the AI ​​model to based on the results of the determination. In this embodiment, an example is described in which a user creates an object detection model for detecting a specific person and intends to publish the model. Hereinafter, a model trained by a user will be referred to as a trained model. However, an object detection AI model for detecting a specific person is merely an example and is not limited to this. Other possible applications include object detection tasks such as races, animals, and vehicles, as well as generation AI for voice recognition and chatbots.

[0022] <Functional configuration of information processing system> FIG. 4 shows a functional block diagram of an information processing system 400 according to this embodiment. The server-side information processing device 100 includes a server receiving unit 401, an evaluation data storage unit 402, a model evaluation unit 403, a disclosure range determination unit 104, a model disclosure unit 405, and a server transmission unit 406. The user-side information processing device 200 includes a model information acquisition unit 409, a user transmission unit 410, a user reception unit 411, and a review result collection unit 412. The user transmission unit 410 and the server reception unit 401, and the server transmission unit 406 and the user reception unit 411 communicate information via a network. However, FIG. 4 is merely an example of a functional configuration and does not limit the scope of application of this embodiment.

[0023] The server receiving unit 401 receives information from the user-side information processing device 200 via a network. The received information is model information of a trained model. The server receiving unit 401, as well as the server transmitting unit 406, user transmitting unit 410, and user receiving unit 411 (described later) convert the data into an appropriate format before transmission, and after reception, convert it into data in a format that is easy to use at the output destination. The model information may include, for example, entity data of the trained model, information such as the target object, the name of the user who trained the model, and final review results from limited users. However, the type of model information acquired is not limited to this, and may be entity data of the model and any information related to the model. The received model information is output to the model evaluation unit 403 and the disclosure range determination unit 404.

[0024] The evaluation data storage unit 402 stores data for calculating the evaluation accuracy r of the trained model regarding AI ethics. The data is divided by evaluation item and comes in various formats, such as a data set of animals that are easily mistaken for humans, a data set of buildings, and a data set on people by race, and each data set is provided with correct answer data.

[0025] The model evaluation unit 403 receives model information, performs inference and evaluation of the trained model included in the model information, and calculates the evaluation accuracy r. The model evaluation unit 403 includes an evaluation data determination unit 407 and an accuracy calculation unit 408. The evaluation data determination unit 407 identifies evaluation items for AI ethics linked to the model information, and selects and determines appropriate evaluation data based on the evaluation items. Examples of evaluation items for AI ethics include false positive detection of humans and non-humans, racial fairness, etc., but they do not need to be limited to one and are not limited to these.

[0026] The accuracy calculation unit 408 calculates the evaluation accuracy r of the trained model using the selected evaluation data. The evaluation accuracy r refers to a comparable evaluation measure such as the value of "100 - false positive rate (percentage)" or the standard deviation of the accuracy rate for each race. The evaluation accuracy r and AI ethics evaluation items are assigned to the model information, and the model information is output to the disclosure range determination unit 404.

[0027] The disclosure range determination unit 404 obtains the difference between the evaluation accuracy r of the model information and a predetermined reference value s associated with each evaluation item, and determines the disclosure range of the trained model based on this difference. Note that the reference value s is set to the maximum value within the range that is acceptable from an AI ethics perspective, for example, with respect to false detection of humans and non-humans, racial fairness, etc. The value of the reference value s is not particularly limited, but may be set to a high value to reduce the risk of AI ethics issues occurring in the market.

[0028] 5A is a block diagram showing the configuration of the disclosure range determination unit 404 according to this embodiment. The disclosure range determination unit 404 includes an accuracy comparison unit 500, a user determination unit 501, and a reference value change unit 502.

[0029] The accuracy comparison unit 500 compares the evaluation accuracy r calculated by the accuracy calculation unit 408 with a reference value s set in advance based on AI ethics for the evaluation item, and obtains the difference. The result of the comparison may be, for example, the difference between the two values, or the ratio of the two values ​​(for example, evaluation accuracy r / reference value s). In this embodiment, the comparison result will be described as a difference between the values.

[0030] The user determination unit 501 determines users to whom the trained model will be made public on the server based on the difference between the accuracy of the trained model acquired by the accuracy comparison unit 500 and the reference value s. Based on the difference, the user determination unit 501 determines whether to make the trained model public to all users, to a limited number of users, or to only users who have trained the trained model. The determined information is then attached to the model information, and the model information is output to the model publication unit 405. Hereinafter, the "limited number of users" will be referred to as "evaluation users." These are users who provide review results in which the trained model has been reviewed (evaluated) from the perspective of AI ethics. The evaluation users can be, for example, approximately 10 to 20 users.

[0031] The reference value changing unit 502 changes the reference value based on the review results of the trained model from the evaluation user in the model information.

[0032] The model publishing unit 405 receives information about users for whom publication has been decided, and outputs a model publication signal to the server transmitting unit 406 so that the model information can be published only to users for whom publication has been decided. The server transmitting unit 406 receives the model publication signal, and outputs it to the user receiving unit 411 of the determined user for transmission to the information processing device 200 on the side of the user for whom publication has been decided based on the signal.

[0033] The model information acquisition unit 409 acquires AI model information that has been learned by the user. The acquired model information is also output to the user transmission unit 410. The user transmission unit 410 receives model information from the model information acquisition unit 409 and outputs it to the server reception unit 401 for transmission to the server-side information processing device 100. The user reception unit 411 receives information from the server-side information processing device 100 via the network. The received information is model information.

[0034] When the trained model is made public to the evaluator user, the review result collection unit 412 acquires the review results in which the evaluator user reviews the trained model from the perspective of AI ethics. Examples of review methods include, but are not limited to, a method of assigning a score out of 5 or 10 points. In this embodiment, the review method is described as a method of assigning a score out of 5 points. The review result collection unit 412 calculates a final review result from the acquired review results and outputs it to the model information acquisition unit 409.

[0035] <Processing> Next, the processing executed by the information processing system 400 according to this embodiment will be described in detail with reference to the flowchart of Fig. 6. The processing content shown in the flowchart of Fig. 6 starts, for example, when a user inputs learned model information. However, the information processing system does not necessarily have to perform all of the steps described in this flowchart.

[0036] In this embodiment, the disclosure scope of the trained AI model is determined based on the evaluation accuracy r of the trained AI model and the reference value s. Specifically, if "evaluation accuracy r ≧ reference value s," the trained AI model is disclosed to all users. If "evaluation accuracy r < reference value s and the difference between them ≦ limit value t," that is, if the evaluation accuracy is slightly below the reference value, the trained AI model is disclosed to evaluator users, which are a portion of users who evaluate the trained AI model. If "evaluation accuracy r < reference value s and the difference between them > limit value t," that is, if the evaluation accuracy is significantly below the reference value, the trained AI model is disclosed only to the user who trained it.

[0037] As a preparation for implementing this flowchart, the user-side information processing device 200 initializes the system. That is, a program is read from the storage medium 202, which is a ROM, and the user-side information processing device 200 is set to an operable state. The server-side information processing device 100 is also set to an operable state in the same way. Then, the user-side information processing device 200 and the server-side information processing device 100 are set to a state in which they can communicate with each other using the user transmitting unit 410, server receiving unit 401, server transmitting unit 406, and user receiving unit 411.

[0038] In S600, the model information acquisition unit 409 acquires the registered model information. The model information acquisition unit 409 performs data conversion on the acquired model information, compressing and encrypting it into a format suitable for transmission. The acquired model information is then output to the user transmission unit 410.

[0039] In S601, user transmitting unit 410 transmits model information received from model information acquisition unit 409 to server receiving unit 401. Server receiving unit 401 decrypts and decompresses the received model information to convert it back to its original format. Then, it outputs the model information to model evaluation unit 403 and disclosure range determination unit 404.

[0040] In S602, the server-side information processing device 100 determines whether the number of times the process of S601 has been performed is 1. If the number of times the process has been performed is 1, the process proceeds to S603. On the other hand, if the number of times the process has been performed is other than 1, the process proceeds to S608.

[0041] In S603, the evaluation data determination unit 407 of the model evaluation unit 403 identifies evaluation items of AI ethics linked to the model information, and selects and determines evaluation data to be used for evaluation from the evaluation data storage unit 402 based on the evaluation items. Hereinafter, the data selected from the evaluation data storage unit 402 will be referred to as the "evaluation set." Furthermore, the accuracy calculation unit 408 performs inference and evaluation of the trained model for the evaluation set, and calculates the evaluation accuracy r. In this embodiment, since the target of the trained model is a specific person, an image dataset for measuring false detection of animals and buildings other than people is selected as the evaluation set.

[0042] The evaluation set also includes image data and frame information indicating the position of the target object within each image. The inference result of the trained model is frame information indicating the position of the object detected as a specific person for each image in the evaluation set. The evaluation accuracy r is a value expressed as "100 - false positive rate (percentage)" when the inference result is compared with the frame information of the evaluation set. In other words, this means that the accuracy is such that no false positives are made for objects other than people. In this example, the evaluation accuracy r is assumed to be 80%. The model evaluation unit 403 adds the obtained evaluation accuracy r and information on the selected evaluation set to the model information and outputs the model information to the disclosure range determination unit 404.

[0043] In S604, the accuracy comparison unit 500 of the disclosure range determination unit 404 compares the evaluation accuracy r with a predetermined reference value s for an evaluation item of AI ethics, and obtains the difference. In this embodiment, the evaluation item of AI ethics is the false positive rate for non-humans, and the reference value s is 90%. The accuracy comparison unit 500 adds the obtained difference information to the model information, and outputs the model information to the user determination unit 501.

[0044] In S605, the user determination unit 501 determines whether to make the trained model public only to the evaluation user. If the result of this step is Yes, the process proceeds to S606. On the other hand, if the result of this step is No, the process proceeds to S610.

[0045] The processing content of S605 will be described in detail below. First, the user determination unit 501 determines the users for whom the trained model will be made available on the server based on the acquired differences. The model publication unit 405 also receives information about the users to be published, and outputs a model publication signal to the server transmission unit 406 so that the model information can be published only to the target users.

[0046] More specifically, if the evaluation accuracy r is equal to or greater than the reference value s, the disclosure range determination unit 404 determines to make the trained model publicly available to all users on the server.

[0047] Furthermore, if the evaluation accuracy r is less than the reference value s and the difference is equal to or less than a preset limit value t, the disclosure range determination unit 404 decides to make the trained model available to evaluator users on the server. Here, the limit value t is set to 20%. Evaluator users are limited to people who can use the trained model and evaluate the AI ​​ethics of the model. They are also limited to people who can be judged to be sufficiently trustworthy based on the user information registered on the server. There is no limit to the number of evaluator users, but the number may be kept to a minimum to reduce the risk of AI ethics issues arising in the market due to the trained model.

[0048] Furthermore, if the evaluation accuracy r is less than the reference value s and the difference exceeds a preset limit value t, the disclosure range determination unit 404 determines to make the trained model available on the server only to users who have trained the trained model. The disclosure range determination unit 404 adds information about the users for which these three patterns have been determined to the model information, and then outputs the model information to the model disclosure unit 405.

[0049] The model publishing unit 405 acquires information about users for whom publication has been decided from the received model information, and outputs a model publication signal and the model information to the server transmission unit 406 so that the model information can be published only to users for whom publication has been decided.

[0050] In S606, the server transmitting unit 406 transmits the model information only to the user receiving unit 411 of the evaluating user based on the model publishing signal received from the model publishing unit 405. The user receiving unit 411 also decrypts and decompresses the received model information to convert it back to its original format. The converted model information can be viewed and downloaded by the evaluating user using the user-side information processing device 200. Here, it is assumed that the user-side information processing device 200 is a PC, and the trained model is downloaded to a single-lens reflex camera via the PC.

[0051] In S607, the review result collection unit 412 acquires the review results of the evaluator user's review of the trained model from the perspective of AI ethics. The evaluator checks the AI ​​ethics evaluation items for which the evaluation accuracy r did not reach the reference value s from the model information, and uses the SLR camera to which the trained model has been downloaded in a scene to which that evaluation item applies. Here, because false detection of non-human subjects is an issue, the evaluator uses the SLR camera for at least a certain period of time or a certain number of shots, for example, at a zoo. After use, the user inputs a review of their impressions of the AI ​​ethics evaluation items into the user's information processing device 200. The review period is not particularly limited, but can be a period sufficient to evaluate any AI ethics issues caused by the trained model. After the review period, the review results of the evaluator users (for example, ratings out of a maximum of 5 stars: ratings ranging from 1 star to 5 stars) are averaged, and this is set as the final review result y. Here, the final review result y is assumed to be 4.5. The final review result y is assigned to the model information, and the model information is output to the model information acquisition unit 409. Then, the process returns to S600.

[0052] The process of S608 is performed when the result of the determination in S602 indicates that the evaluation process is not the first time. In S608, the reference value change unit 502 of the disclosure range determination unit 404 obtains the final review result y from the model information and changes the reference value s. The changed reference value is represented as s'. The reference value s can be changed, for example, according to Equation 1, but is not limited to this and can be changed as appropriate depending on the format of the final review result.

[0053]

number

[0054] In Equation 1, w indicates the weight for adopting the final review result, and g indicates the required review value. The required review value g is the minimum required value when publishing a trained model on the server at the scale of the final review result, and must be set in advance. If the final review result y exceeds the required value g, the standard value s is changed downward by the amount of the excess. If the final review result y does not exceed the required value g, the standard value s is changed upward by the amount of the excess.

[0055] In S609, the accuracy comparison unit 500 of the disclosure range determination unit 404 compares the evaluation accuracy r with the reference value s' changed by the reference value change unit 502, and obtains the difference. In this embodiment, the changed reference value is assumed to be 75%. Information about the obtained difference is added to the model information, and the model information is output to the user determination unit 501.

[0056] In S610, the user determination unit 501 determines the user to whom the trained model will be published on the server based on the difference information added to the acquired model information. The model publishing unit 405 receives the information of the user to be published and outputs a model publication signal to the server transmission unit 406 so that the model information can be published only to that user.

[0057] If the evaluation accuracy r is equal to or greater than the changed reference value s', the user determination unit 501 decides to make the trained model public on the server to all users. If the evaluation accuracy r is less than the changed reference value s', the user determination unit 501 decides to make the trained model public on the server only to users who have trained the trained model. Then, the user determination unit 501 assigns information about the determined users to the model information and outputs the model information to the model publication unit 405. The model publication unit 405 acquires information about users whose information has been decided to be published from the received model information. Then, the model publication unit 405 outputs a model publication signal and the model information to the server transmission unit 406 so that the model information can be published only to users whose information has been decided to be published.

[0058] In S611, the server transmitting unit 406 transmits model information only to the user receiving unit 411 of the user for whom publication was decided in S605 or S610, based on the model publication signal received from the model publishing unit 405. The user receiving unit 411 also decrypts and decompresses the received model information, converting it to its original format. The converted model information can be viewed and downloaded on the user's information processing device 200 by the user for whom publication was decided.

[0059] As described above, in this embodiment, information about the trained AI model is obtained from the user-side information processing device and evaluated, and based on the evaluation results, the server-side information processing device determines the disclosure range to which the trained AI model will be made public.

[0060] According to this embodiment, it is possible to select the range of users to whom a trained model will be made public, reduce the risk of AI ethics issues arising in the market, and then make the trained model public on a server. This prevents trained models from being made publicly available due to excessive caution regarding AI ethics, thereby revitalizing the market while also allowing for consideration of AI ethics.

[0061] (Second embodiment) In this embodiment, an example will be described in which a user determines whether a trained AI model satisfies AI ethics and determines the region (for example, the whole world or some countries or regions) in which to publish the AI ​​model based on the result of the determination. In this embodiment, an example will be described in which a user creates an object detection model for detecting basketball players and intends to publish the model. The system configuration is the same as in the first embodiment, so detailed description will be omitted.

[0062] <Functional configuration> 5(b) is a block diagram showing the configuration of the disclosure range determination unit 404 included in the server-side information processing device 100 according to this embodiment. The disclosure range determination unit 404 according to this embodiment includes an accuracy comparison unit 503, a region determination unit 504, and a reference value change unit 505.

[0063] The function of the precision comparison unit 503 is similar to the function of the precision comparison unit 500 described in the first embodiment, and therefore a description thereof will be omitted.

[0064] The region determination unit 504 determines the region in which the trained model is to be made public on the server based on the difference between the accuracy of the trained model acquired by the accuracy comparison unit 503 and the reference value s. Depending on the difference, the region determination unit 504 determines whether the trained model should be made public to users in all regions around the world, to users in limited regions and some users outside the limited regions, or to only users who have trained the trained model. The region determination unit 504 also adds this information to the model information and outputs the model information to the model publication unit 405. Hereinafter, the some users outside the limited regions will be referred to as evaluation users.

[0065] The function of the reference value changing unit 505 is similar to the function of the reference value changing unit 502 described in the first embodiment, and therefore a description thereof will be omitted.

[0066] The model publication unit 405 in this embodiment receives information about users in the area where publication has been decided, and outputs a model publication signal to the server transmission unit 406 so that the model information can be published only to users in the area where publication has been decided.

[0067] Based on the received model disclosure signal, the server transmission unit 406 in this embodiment outputs it to the user receiving unit 411 of the user in the determined area for transmission to the information processing device 200 on the user side in the area where disclosure has been decided.

[0068] <Processing> Next, the processing executed by the information processing system 400 according to this embodiment will be described in detail with reference to the flowchart of FIG.

[0069] In this embodiment, the disclosure scope of the trained AI model is determined based on the evaluation accuracy r of the trained AI model and a reference value s. Specifically, if the evaluation accuracy r is equal to or greater than the reference value s, it is determined that the trained AI model is to be disclosed to users worldwide. If the evaluation accuracy r is less than the reference value s and the difference between the two is equal to or less than a preset limit value t, it is determined that the trained AI model is to be disclosed to users in a certain region. If the evaluation accuracy r is less than the reference value s and the difference between the two exceeds the preset limit value t, as in the first embodiment, the trained AI model is disclosed only to the user who trained it.

[0070] The processing of S600 to S602 is the same as the processing of S600 to S602 described in the first embodiment, and therefore a description thereof will be omitted.

[0071] In S603, the evaluation data determination unit 407 of the model evaluation unit 403 identifies evaluation items of AI ethics linked to the model information, and selects and determines evaluation data to be used for evaluation from the evaluation data storage unit 402 based on the evaluation items. Hereinafter, the data selected from the evaluation data storage unit 402 will be referred to as the "evaluation set." The accuracy calculation unit 408 performs inference and evaluation of the trained model for the evaluation set, and calculates the evaluation accuracy r. In this embodiment, the target of the trained model is basketball players, and various races are included. Therefore, an image dataset for measuring the detection accuracy of each race is selected as the evaluation set.

[0072] The evaluation set also includes image data and frame information indicating the position of the target object within each image. The inference result of the trained model is frame information indicating the position of the object detected as a basketball player for each image in the evaluation set. When the inference result is compared with the frame information in the evaluation set, the evaluation accuracy r is a value expressed as "100 - [deviation of detection accuracy (percentage) for each race]". This represents the low variation in detection accuracy for each race, that is, racial fairness. In this example, the evaluation accuracy r = 75%. The model evaluation unit 403 adds the obtained evaluation accuracy r and information on the selected evaluation set to the model information and outputs the model information to the disclosure range determination unit 404.

[0073] In S604, the accuracy comparison unit 503 of the disclosure range determination unit 404 compares the evaluation accuracy r with a predetermined reference value s for the evaluation item of AI ethics, and obtains the difference. In this embodiment, the evaluation item of AI ethics is racial fairness, and the reference value s is 90%. The accuracy comparison unit 503 adds the obtained difference information to the model information, and outputs the model information to the region determination unit 504.

[0074] In S605, the region determination unit 504 determines whether to make the trained model public only to the evaluation user. If the result of this step is Yes, proceed to S606. On the other hand, if the result of this step is No, proceed to S610.

[0075] The processing content of S605 will be described in detail below. First, the region determination unit 504 determines the region in which the trained model will be made available on the server based on the acquired difference. Then, the model publication unit 405 receives information about the region in which the model will be published, and outputs a model publication signal to the server transmission unit 406 so that the model information can be published only to users in that region.

[0076] More specifically, if the evaluation accuracy r is equal to or greater than the reference value s, the disclosure range determination unit 404 determines to make the trained model available to users in all regions on the server. Furthermore, if the evaluation accuracy r is less than the reference value s and the difference is equal to or less than a preset limit value t, the disclosure range determination unit 404 compares it with a regulation value v preset for each region. The regulation value v is set for each region based on the AI ​​ethics established for each region. Here, it is assumed that the regulation value v for a certain region is 10%. The limit value t is the lower limit of the regulation values ​​v for all regions where the regulation value v is set, and here it is assumed that the limit value t is 30%.

[0077] As a result of the comparison, the disclosure range determination unit 404 decides to make the trained model available on the server to users in regions where the difference is equal to or less than the regulated value v. Furthermore, the disclosure range determination unit 404 determines that some of the users in regions where the difference exceeds the regulated value v are evaluator users, and decides to make the trained model available on the server to these evaluator users. Evaluator users are limited to people who can use the trained model and evaluate the AI ​​ethics of the model. Furthermore, they are people who can be determined to be sufficiently trustworthy based on the user information registered on the server. There is no limit to the number of evaluator users, but at least one person should be selected for each region, and the number may be kept to a minimum to reduce the risk of AI ethics issues arising in the market due to the trained model.

[0078] Furthermore, if the evaluation accuracy r is less than the reference value s and the difference exceeds the limit value t, the disclosure range determination unit 404 determines to make the trained model available on the server only to users who have trained the trained model. The disclosure range determination unit 404 adds the determined region and user information to the model information and outputs the model information to the model disclosure unit 405.

[0079] The model publishing unit 405 acquires information about the region where publication has been decided from the received model information, and outputs a model publication signal and the model information to the server transmission unit 406 so that the model information can be published only to users in the region where publication has been decided and to rating users.

[0080] The process of S606 is the same as the process of S606 in the first embodiment, and therefore a description thereof will be omitted.

[0081] In S607, the review result collection unit 412 acquires review results in which the evaluator user reviews the trained model from the perspective of AI ethics. The evaluator checks the model information for AI ethics evaluation items for which the evaluation accuracy r does not reach the reference value s, and uses the SLR camera to which the trained model has been downloaded in a scene to which that evaluation item applies. In this case, because racial fairness among basketball players is an issue, the evaluator uses the SLR camera for a certain period of time or a certain number of shots, for example, while watching a sports game where people of various races can be photographed. After use, the user inputs a review of their impressions of the AI ​​ethics evaluation items into the user's information processing device 200. The review period is not particularly limited, but can be set to a period sufficient to evaluate the AI ​​ethics issues caused by the trained model. After the review period, the review results of the evaluators are averaged and set as the final review result y. Here, the final review result y is assumed to be 4.5. The final review result y is assigned to the model information, and the model information is output to the model information acquisition unit 409. Then, the process returns to S600.

[0082] The processes in S608 to S609 are the same as those in S608 to S609 described in the first embodiment, and therefore will not be described again.

[0083] In S610, the region determination unit 504 determines the region in which the trained model is to be made public on the server based on the difference information added to the acquired model information. The model publication unit 405 also receives information on the users to be published and outputs a model publication signal to the server transmission unit 406 so that the model information can be published only to users in the region in which it is to be published.

[0084] If the evaluation accuracy r is equal to or greater than the changed reference value s', the region determination unit 504 decides to make the trained model available on the server to users in all regions of the world. If the evaluation accuracy r is less than the reference value s' and the difference is less than or equal to the limit value t, the region determination unit 504 decides to make the trained model available on the server only to users in regions where the difference is less than or equal to the regulation value v. If the evaluation accuracy r is less than the reference value s' and the difference exceeds the limit value t, the region determination unit 504 decides to make the trained model available on the server only to users who have trained the trained model. Then, the region determination unit 504 assigns information about the determined region to the model information and outputs the model information to the model publication unit 405. The model publication unit 405 acquires information about the region for which publication has been determined from the received model information. Then, the model publication unit 405 outputs a model publication signal and model information to the server transmission unit 406 so that the model information can be published only to users in the region for which publication has been determined.

[0085] In S611, the server transmitting unit 406 transmits model information only to the user receiving units 411 of users in the areas where publication was decided in S605 or S610, based on the model publication signal received from the model publishing unit 405. The user receiving unit 411 also decrypts and decompresses the received model information data to convert it back to its original format. The converted model information can be viewed and downloaded on the user's information processing device 200 by users in the areas where publication was decided.

[0086] As described above, according to this embodiment, it is possible to select the regions in which the trained model will be made public, reduce the risk of AI ethics issues arising in each region, and then make the trained model public on a server.

[0087] (Third embodiment) In this embodiment, an example is described in which a user determines whether a trained AI model satisfies AI ethics, and determines the conditions and user for publishing the AI ​​model based on the determination result. As in the first embodiment, this embodiment is described using an example in which a user creates an object detection model for detecting a specific person and attempts to publish the model. The system configuration is the same as in the first embodiment, so detailed description will be omitted.

[0088] <Functional configuration> 5(c) is a block diagram showing the configuration of the disclosure range determination unit 404 included in the server-side information processing device 100 according to this embodiment. The disclosure range determination unit 404 according to this embodiment includes an accuracy comparison unit 506, an annotation assignment unit 507, a condition determination unit 508, and a reference value change unit 509.

[0089] The function of the precision comparison unit 506 is similar to the function of the precision comparison unit 500 described in the first embodiment, and therefore a description thereof will be omitted.

[0090] The annotation adding unit 507 adds an annotation when publishing the trained model based on the difference between the accuracy of the trained model acquired by the accuracy comparing unit 506 and the reference value s. Details of the annotation content will be described later. The annotation adding unit 507 adds this information to the model information, and then outputs the model information and the information on the difference acquired by the accuracy comparing unit 508 to the condition determining unit 508.

[0091] The condition determination unit 508 determines disclosure conditions for the trained model to which annotations have been added by the annotation adding unit 507 so that the model can be made public only to users who agree with the content of the annotations. The condition determination unit 508 also determines a portion of users from all users who can evaluate the trained model. Hereinafter, these portions of users will be referred to as evaluation users. If no annotation information has been added to the trained model, the condition determination unit 508 determines whether to make the trained model public to all users or only to users who have studied the trained model, based on the difference information. The condition determination unit 508 adds the determined information to the model information and outputs it to the model publication unit 405.

[0092] The function of the reference value changing unit 509 is similar to that of the reference value changing unit 502 described in the first embodiment, and therefore a description thereof will be omitted.

[0093] The model publishing unit 405 receives the model information, and if annotation information has been added to the trained model, outputs a model publishing signal to the server transmitting unit 406 so that the model information can be published only to users who satisfy the publishing conditions. In other cases, the model publishing unit 405 outputs a model publishing signal to the server transmitting unit 406 so that the model information can be published only to users for whom publication has been decided.

[0094] <Processing> Next, the processing executed by the information processing system 400 according to this embodiment will be described in detail with reference to the flowchart of FIG.

[0095] In this embodiment, the disclosure range of the trained AI model is determined based on the evaluation accuracy r of the trained AI model and the reference value s. For example, if the evaluation accuracy r is less than the reference value s and the difference between the two is equal to or less than a preset limit value t, annotation information is added to the trained AI model, and it is determined that the trained AI model will be made public to users who agree to the content of the annotation information.

[0096] The processing of S600 to S603 is the same as the processing of S600 to S603 described in the first embodiment, and therefore a description thereof will be omitted.

[0097] In S604, the accuracy comparison unit 506 of the disclosure range determination unit 404 compares the evaluation accuracy r with a preset reference value s for the evaluation item of AI ethics, and obtains the difference. In this embodiment, the evaluation item of AI ethics is the false positive rate for non-humans, and the reference value s is 90%. The accuracy comparison unit 506 adds the obtained difference information to the model information, and outputs the model information to the annotation adding unit 507.

[0098] In S605, the disclosure range determination unit 404 determines whether or not to disclose the trained model only to the evaluation user. If the result of this step is Yes, proceed to S606. On the other hand, if the result of this step is No, proceed to S610.

[0099] The processing content of S605 will be described in detail below. First, the annotation adding unit 507 adds annotations to the trained model when it is made public on the server based on the acquired differences. Next, the condition determining unit 508 determines the conditions and users when making the trained model public on the server based on the presence or absence of annotations and the differences. Finally, the model publishing unit 405 receives the conditions for publishing and user information, and outputs a model publishing signal to the server transmitting unit 406 so that the model information can be published only to users who satisfy the conditions and users for whom publication has been decided.

[0100] If the evaluation accuracy r is less than the reference value s and the difference is equal to or less than a preset limit value t, the annotation unit 507 adds an annotation to the trained model. Here, the limit value t is set to 20%. Examples of annotation information include, but are not limited to, "Do not use in situations where there are items that are problematic from an AI ethics perspective, so do not use in situations where such items apply" and "Do not publish results of using the above model on social media." Furthermore, if an annotation is added to the trained model, it is decided to make the model public so that only users who read the annotation and agree to its contents can download it.

[0101] In addition, if annotation information is added, evaluation users who will review the model will be selected from all users. Evaluation users will be limited to people who use the trained model and can evaluate the AI ​​ethics of the model. They will also be people who can be judged to be sufficiently trustworthy based on the user information registered on the server. There is no limit to the number of evaluation users, but the number may be kept to a minimum to reduce the risk of AI ethics issues arising in the market due to trained models.

[0102] Furthermore, if the evaluation accuracy r is equal to or greater than the reference value s, the disclosure range determination unit 404 decides to make the trained model available on the server to all users without any conditions. If the evaluation accuracy r is less than the reference value s and the difference exceeds a predetermined limit value t, the disclosure range determination unit 404 decides to make the trained model available on the server without any conditions only to users who have trained the trained model. The disclosure range determination unit 404 adds the determined information to the model information and outputs the model information to the model disclosure unit 405.

[0103] The model publishing unit 405 acquires the conditions for publication and information about users for whom publication has been decided from the received model information. The model publishing unit 405 outputs a model publication signal and the model information to the server transmission unit 406 so that the model information can be published only to users who meet the conditions for publication and users for whom publication has been decided.

[0104] The processes in S606 to S609 are the same as those in S606 to S609 in the first embodiment, and therefore the description thereof will be omitted.

[0105] In S610, the annotation unit 507 adds annotations to the trained model when it is made public on the server based on the acquired differences. Next, the condition determination unit 508 determines the conditions and users for making the trained model public on the server based on the presence or absence of annotations and the differences. Finally, the model publication unit 405 receives the conditions for publication and user information, and outputs a model publication signal to the server transmission unit 406 so that the model information can be published only to users who satisfy the conditions and users for whom publication has been decided.

[0106] If the evaluation accuracy r is less than the reference value s' and the difference is equal to or less than a preset limit value t, the annotation unit 507 adds an annotation to the trained model. Furthermore, if an annotation has been added to the trained model, the annotation unit 507 decides to make the model public so that only users who read the annotation and agree to its contents can download it.

[0107] If the evaluation accuracy r is equal to or greater than the reference value s', the annotation unit 507 decides to make the trained model available to all users on the server without any conditions. If the evaluation accuracy r is less than the reference value s' and the difference exceeds a preset limit value t, the model publication unit 405 decides to make the trained model available to only the user who learned the trained model on the server without any conditions.

[0108] The annotation adding unit 507 adds the above-mentioned determined information to the model information, and outputs the model information to the model publishing unit 405. The model publishing unit 405 acquires the conditions for publication and information on users for whom publication has been decided from the received model information. The model publishing unit 405 outputs a model publication signal and the model information to the server transmission unit 406 so that the model information can be published only to users who meet the conditions for publication and users for whom publication has been decided.

[0109] In S611, the server transmitting unit 406 receives the model publication signal from the model publishing unit 405, and transmits model information only to the user receiving units 411 of users who satisfied the conditions in S605 and S610 and users for whom publication has been decided. The user receiving unit 411 also decrypts and decompresses the received model information to convert it back to its original format. The converted model information can be viewed and downloaded on the user's information processing device 200 by users who satisfied the conditions and users for whom publication has been decided.

[0110] As described above, according to this embodiment, when a trained model that has AI ethics issues is made public, annotations can be added and the model can be made public to users who agree to the annotations. By allowing the users to use the trained model after understanding the AI ​​ethics issues of the trained model from the annotations, the risk of AI ethics issues arising in the market can be reduced.

[0111] The disclosure of this specification includes the following information processing device, control method for an information processing device, and program.

[0112] (Item 1) An information processing device capable of communicating with other information processing devices, An acquisition means for acquiring information about a trained AI model from the other information processing device; An evaluation means for evaluating the trained AI model; A determination means for determining a disclosure range in which the information processing device will disclose the trained AI model based on the evaluation result of the evaluation means; An information processing device comprising:

[0113] (Item 2) The evaluation means A determination means for identifying evaluation items for AI ethics based on information about the trained AI model and determining evaluation data according to the evaluation items; a calculation means for calculating the evaluation accuracy of the trained AI model based on the evaluation data; 2. The information processing device according to item 1, comprising:

[0114] (Item 3) Item 3. The information processing device according to item 2, wherein the determination means determines the disclosure range of the trained AI model based on the evaluation accuracy of the trained AI model and a reference value associated with the evaluation item.

[0115] (Item 4) 4. The information processing device according to item 3, wherein the decision means decides to make the trained AI model public to all users if the evaluation accuracy is equal to or greater than the reference value.

[0116] (Item 5) 5. The information processing device according to item 3 or 4, wherein the decision means decides to make the trained AI model public to evaluation users, who are a portion of users who evaluate the trained AI model, when the evaluation accuracy is less than the reference value and the difference between the evaluation accuracy and the reference value is equal to or less than a preset limit value.

[0117] (Item 6) The information processing device described in any one of items 3 to 5, wherein the decision means decides to make the trained AI model public to a user who trained the trained AI model when the evaluation accuracy is less than the reference value and the difference between the evaluation accuracy and the reference value exceeds a preset limit value.

[0118] (Item 7) 4. The information processing device according to item 3, wherein the decision means decides to make the trained AI model public to the entire world if the evaluation accuracy is equal to or greater than the reference value.

[0119] (Item 8) 8. The information processing device according to item 3 or 7, wherein the decision means decides to make the trained AI model public to a certain region when the evaluation accuracy is less than the reference value and the difference between the evaluation accuracy and the reference value is equal to or less than a preset limit value.

[0120] (Item 9) Item 4. The information processing device according to item 3, wherein the decision means, when the evaluation accuracy is less than the reference value and the difference between the evaluation accuracy and the reference value is equal to or less than a preset limit value, adds annotation information to the trained AI model and decides to make the trained AI model public to users who agree to the content of the annotation information.

[0121] (Item 10) an acquisition means for acquiring a review result of the trained AI model by the evaluation user; a change means for changing the reference value based on the review result; 6. The information processing device according to item 5, further comprising:

[0122] (Item 11) 11. The information processing device according to any one of items 1 to 10, further comprising a publishing means for publishing the trained AI model based on the publishing range.

[0123] (Item 12) A method for controlling an information processing device capable of communicating with another information processing device, an acquisition step of acquiring information about the trained AI model from the other information processing device; an evaluation step of evaluating the trained AI model; a determination step in which the information processing device determines a disclosure range in which the trained AI model will be disclosed based on the evaluation result in the evaluation step; 1. A method for controlling an information processing device, comprising:

[0124] (Item 13) A program for causing a computer to function as the information processing device according to any one of items 1 to 11.

[0125] (Other embodiments) The present invention can also be realized by supplying a program that realizes one or more functions of the above-described embodiments to a system or device via a network or a storage medium, and having one or more processors in the computer of the system or device read and execute the program.The present invention can also be realized by a circuit (e.g., ASIC) that realizes one or more functions.

[0126] The invention is not limited to the above-described embodiments, and various changes and modifications can be made without departing from the spirit and scope of the invention. Accordingly, the following claims are appended to apprise the public of the scope of the invention. [Explanation of symbols]

[0127] 400: Information processing system, 100: Server-side information processing device, 200: User-side information processing device, 401: Server receiving unit, 402: Evaluation data storage unit, 403: Model evaluation unit, 404: Disclosure decision unit, 405: Model disclosure unit, 406: Server transmitting unit, 407: Evaluation data determination unit, 408: Accuracy calculation unit, 409: Model information acquisition unit, 410: User transmitting unit, 411: User receiving unit, 412: Review result collection unit

Claims

1. An information processing device capable of communicating with other information processing devices, An acquisition means for acquiring information on a trained AI model from the other information processing device; A determination means for identifying evaluation items of AI ethics based on information of the trained AI model and determining evaluation data according to the evaluation items; A calculation means for calculating the evaluation accuracy of the trained AI model based on the evaluation data; a determination means for determining a disclosure range in which the information processing device will disclose the trained AI model based on the evaluation accuracy of the trained AI model and a reference value associated with the evaluation item; When the evaluation accuracy is less than the reference value and the difference between the evaluation accuracy and the reference value is equal to or less than a preset limit value, the decision means decides to make the trained AI model public to evaluation users, which are a portion of users who evaluate the trained AI model; The information processing device includes: An acquisition means for acquiring a review result of the trained AI model by the evaluation user; a change means for changing the reference value based on the review result; The information processing device further comprises:

2. The information processing device according to claim 1 , wherein the decision means decides to make the trained AI model public to all users if the evaluation accuracy is equal to or greater than the reference value.

3. The information processing device according to claim 1, characterized in that the decision means decides to make the trained AI model public to a user who trained the trained AI model when the evaluation accuracy is less than the reference value and the difference between the evaluation accuracy and the reference value exceeds the limit value.

4. The information processing device according to claim 1 , wherein the decision means decides to make the trained AI model public to the entire world if the evaluation accuracy is equal to or greater than the reference value.

5. The information processing device according to claim 1, characterized in that the decision means decides to make the trained AI model public to a certain region when the evaluation accuracy is less than the reference value and the difference between the evaluation accuracy and the reference value is equal to or less than the limit value.

6. The information processing device according to claim 1, characterized in that, when the evaluation accuracy is less than the reference value and the difference between the evaluation accuracy and the reference value is equal to or less than the limit value, the decision means adds annotation information to the trained AI model and decides to make the trained AI model public to users who agree to the content of the annotation information.

7. The information processing device according to claim 1 , further comprising a publishing means for publishing the trained AI model based on the publishing range.

8. A method for controlling an information processing device capable of communicating with another information processing device, An acquisition step of acquiring information about the trained AI model from the other information processing device; A determination process of identifying evaluation items of AI ethics based on information of the trained AI model and determining evaluation data according to the evaluation items; A calculation step of calculating the evaluation accuracy of the trained AI model based on the evaluation data; a determination step of determining a disclosure range in which the information processing device will disclose the trained AI model based on the evaluation accuracy of the trained AI model and a reference value associated with the evaluation item, In the determination step, if the evaluation accuracy is less than the reference value and the difference between the evaluation accuracy and the reference value is equal to or less than a preset limit value, it is determined to make the trained AI model public to evaluation users, which are a portion of users who evaluate the trained AI model; The control method includes: an acquisition step of acquiring a review result of the trained AI model by the evaluation user; a modifying step of modifying the reference value based on the review result; 4. The method for controlling an information processing device, further comprising:

9. A program for causing a computer to function as the information processing device according to any one of claims 1 to 7.

Images