Information processing apparatus and control method thereof

Abstract

To control disclosure of a learning model.SOLUTION: An information processing apparatus for managing a plurality of learning models disclosed in a model disclosure platform comprises: management means which, with respect to each of the plurality of learning models, manages first information about a learning history of the learning model and second information about a base model used for generation of the learning model; and determination means which determines a disclosure range of each of the plurality of learning models in the model disclosure platform on the basis of both of the first information and the second information.SELECTED DRAWING: Figure 7

Description

[Technical Field]

[0001] The present invention relates to a management technique for a learning model and learning data. [Background technology]

[0002] In recent years, artificial intelligence (AI) technology using machine learning has made remarkable progress and is now being applied to a variety of products and services, including image recognition and machine translation. Recently, generative AI, which generates content such as text and images based on user input, has also emerged and is attracting attention.

[0003] Many of these AI technologies are realized by learning models (trained models) created by companies, research institutions, etc. However, in the future, it is expected that it will become common for general users to perform machine learning to suit their own purposes and create learning models. In addition, it is expected that general users will be able to perform machine learning and publish learning models via internet-based services. Here, we refer to such services as "learning model publishing platforms."

[0004] On a learning model publishing platform, users can download and use learning models published by other users. It is anticipated that a portion of the download fee will be paid to the model creator as a reward, encouraging users to create models on the platform. When creating a model, users may use a learning model published by another user as an initial model, conduct additional training to suit their own needs, and then publish the resulting learning model. However, published learning models and data used in training must comply with ethics and not infringe copyright or portrait rights.

[0005] Patent Document 1 discloses a technology that enables the creation of a learning model that complies with AI ethics by conducting an AI ethics judgment based on the learning conditions at the start of machine learning and only carrying out learning if it is determined that the AI ​​complies with ethics. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Japanese Patent Application Publication No. 2023-64636 Summary of the Invention [Problem to be solved by the invention]

[0007] However, the AI ​​ethics rules that learning models must comply with may be revised. With the technology described in Patent Document 1, even if the learning model complies with the rules at the time of learning, it is not guaranteed that a learning model created after the rules are revised will comply with the AI ​​ethics rules after the rules are revised.

[0008] Therefore, problems may arise even after a learning model created according to the technology described in Patent Document 1 is made public on a learning model publishing platform. That is, in addition to the published learning model, derived models created by additional learning using the learning model as an initial model may not satisfy the AI ​​ethics standards after the revised regulations. Furthermore, on learning model publishing platforms, there is a risk that derived models with insufficient additional learning (derived models that are substantially indistinguishable from the initial model) may be published. Therefore, managing and controlling not only the publication of learning models but also the publication of derived models has become an issue.

[0009] The present invention has been made in consideration of such problems, and aims to provide a technology for appropriately controlling the publication of learning models. [Means for solving the problem]

[0010] In order to solve the above-mentioned problems, an information processing device according to the present invention has the following configuration. That is, the information processing device manages a plurality of learning models published on a model publishing platform, a management means for managing, for each of the plurality of learning models, first information on the learning history of the learning model and second information on the base model used to create the learning model; a determination means for determining a disclosure range of each of the plurality of learning models on the model disclosure platform based on both the first information and the second information; Equipped with. [Effects of the Invention]

[0011] According to the present invention, a technology for appropriately controlling the publication of a learning model can be provided. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 2 is a block diagram showing a functional configuration of the information processing device. [Figure 2] FIG. 10 is a diagram illustrating an example of information assigned to a model. [Figure 3] FIG. 10 is a diagram illustrating an example of learning data. [Figure 4] FIG. 10 is a diagram illustrating an example of information added to data. [Figure 5] FIG. 10 is a diagram illustrating an example of traceability information. [Figure 6] FIG. 10 is a diagram illustrating an example of disclosure criteria information. [Figure 7] 4 is a flowchart of processing in an information processing device (first embodiment). [Figure 8] 10 is a flowchart of processing in an information processing device (second embodiment). [Figure 9] 10 is a flowchart of processing in an information processing device (third embodiment). [Figure 10] 10 is a flowchart of processing in an information processing device (modification example). [Figure 11] FIG. 10 is a diagram illustrating an example of a traceability information management screen. DETAILED DESCRIPTION OF THE INVENTION

[0013] 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 invention claimed. 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.

[0014] (First embodiment) As a first embodiment of an information processing device according to the present invention, an information processing device that controls the disclosure range of a learning model will be described below. In particular, an example will be described in which the disclosure range of a learning model is narrowed (limited disclosure) or prohibited (private disclosure) when the learning data contains an image that is ethically or legally (copyright, portrait right) inappropriate.

[0015] <Summary> The information processing device of this embodiment manages information (learning data ID) about data used in learning a learning model (trained model) in association with the learning model. Furthermore, if the learning model is a derived model, information (initial model ID) about the base model (initial model) used is managed in association with the learning model. The information processing device then manages traceability information (past history of the learning model) based on the initial model ID and learning data ID of each of the multiple learning models, and controls the disclosure range of each learning model based on the traceability information.

[0016] <Terminology> "Ethically flawed images" are, for example, images that suggest violence or discrimination, or images that contain sexual depictions. "Copyright- and portrait-right-deficient images" are images that infringe copyright or portrait rights. For example, they are images that the rights holder does not permit for widespread public use, such as copying or machine learning. Examples of such images include photos of affiliated entertainers released by talent agencies and images of animated characters. In the following explanation, "ethically flawed images" and "copyright- and portrait-right-deficient images" may be collectively referred to as "NG images."

[0017] <Device configuration> 1 is a block diagram showing the functional configuration of an information processing device 11. The information processing device 11 manages "models" and "data" used for learning and evaluating the models. The information processing device 11 also learns models and edits data in response to requests from users (users of the learning model publishing platform), and edits model information (determines whether to publish) in response to requests from administrators (administrators of the learning model publishing platform).

[0018] The information processing device 11 is communicably connected to the administrator terminal 13 and the user terminal 14 via a communication network such as a local area network (LAN) or the Internet. The method of connection is not particularly limited. For example, they may be connected separately via a wired connection or via wireless communication. Also, while FIG. 1 shows one administrator terminal and one user terminal, this is not limiting. The administrator terminal 13 and the user terminal 14 are assumed to be information terminals such as personal computers (PCs), mobile phones, and tablet terminal devices.

[0019] The information processing device 11 includes a control unit 10 and a storage unit 12. The control unit 10 can be realized, for example, by a central processing unit (CPU) executing various programs stored in the storage unit 12. The storage unit 12 is a large-capacity storage device such as a random access memory (RAM), a read-only memory (ROM), or a hard disk drive (HDD) or a solid-state drive (SSD).

[0020] The control unit 10 includes a model management unit 101, a data management unit 102, a traceability information management unit 103, a disclosure standard information output unit 104, a disclosure determination unit 105, a registration unit 106, a learning and evaluation unit 107, a display control unit 108, and a communication control unit 109. As described above, each of these functional units can be realized by, for example, a CPU executing various programs. However, some or all of these functional units may also be realized by hardware such as an application specific integrated circuit (ASIC).

[0021] The model management unit 101 manages models registered in the information processing device 11. In this embodiment, the model is assumed to be an "object detection model" intended to detect objects from images, but is not limited to this. It may also be a voice recognition model, a natural language processing model, or a generative artificial intelligence (AI) model.

[0022] 2 is a diagram showing examples of model information 201, 202 assigned to two models. The model information is attribute information of the model, and includes information such as a model ID, a user ID, an initial model ID, a learning data ID, tags, tasks, learning parameters, evaluation data IDs, evaluation results, and disclosure range. In other words, it includes the learning history, such as the learning data ID and learning parameters.

[0023] The model ID is a model-specific identification information (ID) for identifying each model. The user ID is a user-specific ID for identifying the user who performed training for each model. The initial model ID is a model ID that specifies the model that became the initial parameter when training each model.

[0024] 2, in model information 201, the model ID is "m0001" and the initial model ID is "m0000". In addition, in model information 202, the model ID is "m0002" and the initial model ID is "m0001". This indicates that the model with the model ID "m0002" is a model obtained as a result of learning using the model with the model ID "m0001" as the initial model.

[0025] The training data ID is a unique ID for each training data set used to train each model. The tag contains information about the objects targeted by each model and information necessary for users to search for models. The task contains the type of processing targeted by each model (object detection, image generation, etc.). The training parameters are the hyperparameters (learning rate, number of training attempts, etc.) used when training each model. The evaluation data ID is a data ID that specifies the dataset used when evaluating each model. The evaluation results contain numerical values ​​obtained as a result of the evaluation. The visibility contains information indicating the visibility of each model (the range of visibility permitted). For example, it can be all (visible to everyone (fully visible)), private, or one or more user IDs. If a user ID is listed, only users with the listed ID will be able to access the model; other users will not be able to access the model.

[0026] 2, model information 201 and 202 are both model information for an object detection model. For example, if the model is another type of model (such as a voice recognition model), the items included in the model information may be different from those in the example of FIG.

[0027] The data management unit 102 manages data registered in the information processing device 11. The data in this embodiment is image data for object detection.

[0028] FIG. 3 is a diagram showing an example of training data. The training data includes an image 301 and a true value 304. The image 301 is an image showing subjects that the user wants to recognize (here, a person 302 and a dog 303). Furthermore, in order to train object detection, not only an image but also subject information as a true value is required. The subject information is, for example, bounding box (BB) information corresponding to information on the position and size of the subject in the image. BB305 is a BB corresponding to the person 302, and BB306 is a BB corresponding to the dog 303.

[0029] 4 is a diagram showing examples of data information 401, 402 assigned to data (image data set). The data information is attribute information of the data, and includes information on the data ID, user ID, tag, task, and image ID.

[0030] The data ID is a data-specific ID used to identify each piece of data. The user ID is a user-specific ID used to identify the user who created each piece of data. The tag contains information about the object that each piece of data targets, as well as information necessary for users to search for data. The task contains the type of processing that each piece of data targets (object detection, image generation, etc.). The image ID is an image-specific ID used to identify one or more images included in the data. Each piece of data contains one or more image IDs.

[0031] 4, data information 401 and 402 are both data information of data used in a model for an object detection model. For example, if the data is used in another type of model (such as a voice recognition model), the items included in the data information may be different from those in the example of FIG.

[0032] The data management unit 102 accepts data registration from the user and registers a new data set.

[0033] The traceability information management unit 103 manages the traceability information of models. Here, the traceability information is information indicating the process by which each model was created in the past (i.e., the past history of the model), and is, for example, information generated based on model information.

[0034] FIG. 5 is a diagram showing an example of traceability information. In FIG. 5, arrows indicate the derivation relationships of each model. For example, model 502 (model ID is "m0002") is a derived model created by additional learning using model 501 (model ID is "m0001") as the initial model. Similarly, model 503 is a derived model of model 502, and model 504 is a derived model of model 501. Note that in FIG. 5, the initial model ID shown in model 501 is shown as "-", which indicates that the initial model is an unlearned model.

[0035] In this way, since the traceability information shows a record of how each model has been edited in the past, it is possible to identify one or more models used to create that model by tracing the traceability information. For example, from the traceability information shown in Figure 5, it can be said that models 502 to 504 are all derived models of model 501. Furthermore, the data used in learning each model can be identified from the learning data ID of each model in Figure 5, and the linking relationship between the model and the data can be shown as the dotted line in Figure 5.

[0036] The disclosure criteria information output unit 104 outputs disclosure criteria information used as one of the criteria for determining whether a model is to be made public. As will be described in detail later, the disclosure criteria information is information that is output for each model based on the model information acquired from the model management unit 101 and the data information acquired from the data management unit 102.

[0037] FIG. 6 is a diagram showing an example of disclosure criteria information. For example, the disclosure criteria information is information indicating whether each model uses a specific image (an image having an image ID specified by the administrator terminal 13) as learning data. For example, assume that an image having an image ID ("imgAAAA1002") shown in bold in FIG. 5 is an "NG image," and the administrator has specified the image as an unusable image via the administrator terminal 13. From FIG. 5, the data IDs of the data containing the image are "d0002" and "d0004," and the models using these data are model 502 and model 504. Therefore, as shown in FIG. 6 (center column), the disclosure criteria information is "FALSE" for model 501 and model 503, and "TRUE" for model 502 and model 504.

[0038] The disclosure determination unit 105 determines whether or not to disclose a model based on the disclosure standard information (first determination result) output by the disclosure standard information output unit 104 and the determination result (second determination result) based on the traceability information managed by the traceability information management unit 103. Note that, as shown in FIG. 6 (right column), the determination result based on the traceability information indicates whether or not each model is a derived model of a model trained using an "NG image." For example, referring to FIG. 5, model 503 is a derived model of model 502, and therefore the determination result is "TRUE."

[0039] The disclosure determination unit 105 then determines whether the model should be made public based on the results of the determination using the disclosure standard information and the traceability information. Here, models 502 and 504 whose disclosure standard information is "TRUE" are determined to be "private," and model 503 whose disclosure standard information is "FALSE" but whose traceability information is "TRUE" is determined to be "limitedly public (public to one or more specific user IDs)." Models that do not fall into any of these categories are determined to be "publicly public."

[0040] The registration unit 106 updates the value of the disclosure range in the model information of each model registered in the information processing device 11 based on the determination result (disclosure, non-disclosure, restricted disclosure) of the disclosure determination unit 105.

[0041] The learning and evaluation unit 107 performs learning and evaluation of the model based on the model and data registered in the information processing device 11 .

[0042] The display control unit 108 controls the display of results on the information processing device 11 in response to requests from the administrator terminal 13 and the user terminal 14. For example, it also generates a management screen (described in the fourth embodiment) to be displayed on the administrator terminal 13. The communication control unit 109 controls the transmission and reception of information between the administrator terminal 13 and the user terminals.

[0043] The storage unit 12 stores various programs executed by the CPU described above, as well as models, model information, data, data information, traceability information, and the like.

[0044] <Device Operation> Fig. 7 is a flowchart of processing in the information processing device 11. Fig. 7(a) is a flowchart when learning and registering a model, and Fig. 7(b) is a flowchart when determining / updating the disclosure range of the model.

[0045] First, with reference to Fig. 7(a), a description will be given of the processing when a user operates the user terminal 14 to register data in the information processing device 11 and learn a model. Here, a description will be given of a situation in which new data (data set) is registered and one of the models already registered is additionally learned using the data.

[0046] In S701, when the registration unit 106 receives a data addition request from the user terminal 14, it registers the data in the storage unit 12. The registered data is assigned data information such as that shown in the data information 401. When registering data, information such as the disclosure range of the data, tags, and tasks can be assigned. Here, the data ID of the data registered by the user is assumed to be "d0001".

[0047] In S702, when the learning / evaluation unit 107 receives a learning execution request from the user terminal 14, it executes learning using data registered by the user. For example, it accepts a model (model ID is "m0000") specification from the user and executes learning using data (data ID is "d0001") for that model. It is assumed that the model with model ID "m0000" is an unlearned model and has been registered in advance in the storage unit 12.

[0048] As a learning method for object detection, for example, a method using a neural network (NN) exists. The learning method for object detection using a NN is described in detail in Literature A. The learning and evaluation unit 107 may evaluate the model trained using the data used for learning or any data.

[0049] (Reference A) Tian et al., "FCOS: Fully Convolutional One-Stage Object Detection", arXiv:1904.01355, 2019 In S703, the registration unit 106 registers the model learned by the user in the storage unit 12. When registering the model, for example, model information such as that shown in model information 201 is assigned. The user ID is the ID of the user operating the user terminal 14.

[0050] In S704, the traceability information management unit 103 updates the traceability information. That is, as shown in Fig. 5, the model 501 (model ID is "m0001") is managed in association with the data (data ID is "d0001").

[0051] As a result, the data prepared by the user (data ID is "d0001") and the model 501 trained using the data (model ID is "m0001") are registered in the information processing device 11.

[0052] Next, we will explain the process for determining / updating the model's disclosure range with reference to Figure 7(b). Here, we will explain the situation when the administrator specifies (registers) a new NG image.

[0053] In S711, upon receiving a request to register an NG image from the administrator terminal 13, the registration unit 106 registers the NG image. For example, the image ID of the NG image may be received directly, or multiple images may be displayed on a display unit (not shown) of the administrator terminal 13, and the administrator may select an NG image. In this example, "imgAAAA1002" shown in Fig. 5 is received and registered as the image ID of the NG image.

[0054] In S712, the data management unit 102 acquires the data IDs of the data including the image IDs of the NG images. In the situation of Fig. 5, "d0002" and "d0004" are acquired.

[0055] In S713, the model management unit 101 acquires a model ID having the data ID acquired in S712 as the learning data ID. In the situation of Fig. 5, "m0002" and "m0004" are acquired.

[0056] In S714, the traceability information management unit 103 acquires the traceability information shown in FIG.

[0057] In S715, the public reference information output unit 104 outputs public reference information for each model ID acquired in S713. Here, "TRUE" is output for models 502 and 504 corresponding to the model ID acquired in S713, and "FALSE" is output for other models (FIG. 6 (center column)).

[0058] In S716, the disclosure determination unit 105 performs disclosure determination for the models stored in the storage unit 12 based on the disclosure standard information output in S715 and the traceability information acquired in S714. First, based on the traceability information acquired in S714, it determines whether each model is a derived model of a model trained using an "NG image" (derived model information). Here, it outputs "TRUE" for model 503, which is the currently focused learning model, and outputs "FALSE" for other models (Figure 6 (right column)). Then, it performs disclosure determination based on both of the two pieces of information for each model (disclosure standard information and derived model information).

[0059] In S717, the registration unit 106 updates the value of the disclosure range in the model information of each model based on the result of the disclosure determination in S716.

[0060] As described above, according to the first embodiment, it is possible to effectively determine the range of disclosure of a model that uses NG images for learning and its derived models.

[0061] (Variation 1-1) The above-mentioned disclosure determination unit 105 determines that a model whose disclosure standard information is "FALSE" and whose derived model information is "TRUE" is to be made available for limited disclosure. However, before determining that a model is to be made available for limited disclosure, model performance verification may be performed on the model, and the result may be used as the final disclosure determination result.

[0062] Here, model performance verification is verification of whether each model has the performance specified by the administrator. For example, if the model is a personal authentication model, verification is performed to verify whether the model has the performance to identify a person (e.g., a person included in the above-mentioned NG image) appearing in an image specified by the administrator terminal 13.

[0063] The disclosure determination unit 105 then sets models whose model performance verification results are "TRUE" (=having the ability to identify a person) as "private" and sets models whose model performance verification results are "FALSE" as "public" or "limited public."

[0064] (Second embodiment) In the second embodiment, another information processing device that controls the disclosure range of a learning model will be described below as an example. In particular, an example of setting the disclosure range to "limited disclosure" or "privacy" based on a generated image obtained as a result of executing the model will be described.

[0065] <Device configuration> The functional configuration of the information processing device according to the second embodiment is the same as that of the first embodiment (FIG. 1), and therefore a detailed description thereof will be omitted. The following description will mainly focus on functional units whose operations differ from those of the first embodiment. In addition, in the following description, the model will be referred to as an "image generation model" that generates illustrations and photographs based on input text information, but the present invention is not limited to this.

[0066] The public reference information output by the public reference information output unit 104 according to the second embodiment is the similarity between a correct answer image and a generated image for a given keyword (for example, the name of an animation character that is unusable due to copyright). Here, the generated image is an image generated when a model executes a keyword input from the administrator terminal 13. Note that a character string of a watermark image (such as a logo or copyright notice) embedded in an image may be used as the keyword.

[0067] The disclosure determination unit 105 performs disclosure determination for the models registered in the storage unit 12 based on the disclosure standard information output by the disclosure standard information output unit 104 and the traceability information managed by the traceability information management unit 103.

[0068] As described above, the disclosure reference information is the similarity between the correct image and the generated image. Therefore, the correct image indicated by the keyword input from the administrator terminal 13 is prepared and registered in advance. For example, if the input keyword is "character A," the correct image of character A is prepared and registered in advance.

[0069] The similarity calculation unit included in the disclosure determination unit 105 then calculates the similarity between the correct image and the generated image, and if the similarity exceeds a predetermined threshold, determines the similarity as "TRUE." A similarity determination of "TRUE" indicates a high possibility that an image that is not permitted for copyright reasons is used in the learning data (although it is not identical to the registered NG image). Therefore, the disclosure determination unit 105 determines that a model for which the similarity determination is "TRUE" and its derivative models are "private" or "limited disclosure." Models that do not fall into either of these categories are determined to be "publicly disclosed."

[0070] Here, the similarity between the correct image for the input keyword and the generated image is used as the judgment condition, but this is not limiting. Also, a classification unit may be provided in the disclosure judgment unit 105, and the judgment condition may be whether the classification result obtained by processing the generated image with a classifier matches the keyword.

[0071] <Device Operation> Among the processes of the information processing device according to the second embodiment, the process of learning and registering a model is the same as that of the first embodiment (FIG. 7(a)), and therefore a description thereof will be omitted.

[0072] FIG. 8 is a flowchart for determining / updating the disclosure range of a model in the second embodiment.

[0073] In S801, when the disclosure criteria information output unit 104 receives a request to register a keyword from the administrator terminal 13, it registers the keyword.

[0074] In S802, when the disclosure determination unit 105 receives a request to register a correct image corresponding to the keyword registered in S801 from the administrator terminal 13, the disclosure determination unit 105 registers the correct image in association with the keyword.

[0075] In S803, the traceability information management unit 103 acquires traceability information. Here, the traceability information acquired is the information shown in Fig. 5. Note that in the second embodiment, the configuration may be such that only derivation information between models (arrows in Fig. 5) is acquired.

[0076] In S804, the public reference information output unit 104 generates an image by inputting the keyword registered in S801 into the public model stored in the storage unit 12. Then, the similarity between the correct image for the keyword and the generated image is output as public reference information.

[0077] In S805, the disclosure determination unit 105 performs disclosure determination of the model stored in the storage unit 12 based on the disclosure standard information output in S804 and the traceability information acquired in S803. The detailed operation of the disclosure determination by the disclosure determination unit 105 is as described above.

[0078] In S806, the registration unit 106 updates the value of the disclosure range in the model information of each model based on the result of the disclosure determination in S805.

[0079] As described above, according to the second embodiment, the similarity between a generated image obtained by inputting a keyword related to an NG image into a model and the correct image for that keyword is calculated, and the disclosure range of that model and its derivative models is determined based on the similarity. In particular, if the similarity exceeds a predetermined threshold, it is determined that there is a high possibility that an image similar to the NG image was used in learning the model, and that model and its derivative models are set to "private" or "limited disclosure." This makes it possible to effectively determine the disclosure range of not only the registered NG image but also models and their derivative models that use images similar to the NG image.

[0080] (Third embodiment) The third embodiment will be described below by taking another information processing device that controls the disclosure range of a learning model as an example.

[0081] <Summary> One of the challenges of model publishing on learning model publishing platforms is that derivative models with substantially the same performance as highly rated models may be registered on the platform. In such cases, the reward that should be paid to the user who created the highly rated model ends up being paid to a third party who registered the derivative model. Therefore, it is necessary to ensure a sense of fairness and satisfaction for the user who created the highly rated model.

[0082] Therefore, when a derived model (also called a child model) is newly registered in the information processing device of this embodiment, the information processing device checks the similarity between the derived model and the initial model (also called a parent model). If the similarity between the derived model and the initial model is higher than a predetermined threshold (if the derived model is substantially identical to the initial model), the information processing device sets the derived model to "limited release" or "private."

[0083] <Device configuration> The functional configuration of the information processing device according to the third embodiment is the same as that of the first embodiment (FIG. 1), and therefore a detailed description thereof will be omitted. The following description will mainly focus on functional units whose operations differ from those of the first embodiment. In addition, in the following description, the model will be referred to as an "object detection model" as in the first embodiment, but this is not intended to be limiting.

[0084] The model management unit 101 of the third embodiment acquires, from the storage unit 12, model information of a model (child model) registered by a user on a user terminal 14 and its parent model, based on the traceability information managed by the traceability information management unit 103. The acquired model information is passed to the disclosure standard information output unit 104.

[0085] The public reference information output unit 104 acquires the learning history (learning data ID, learning parameters) from the model information of each of the parent model and the child model. Then, it determines whether the learning data and learning parameters are the same between both models, and outputs the "determination result" and the "number of times the child model has learned" as public reference information.

[0086] The identity of the training data between the two models is determined by determining the identity / similarity of the training data used in the training. For example, this may be determined based on whether the training data IDs are the same. Alternatively, the identity / similarity of multiple image IDs included in the training data of each of the two models may be determined. For example, if multiple image IDs included in the training data of the child model overlap with multiple image IDs included in the training data of the parent model by a predetermined percentage or more (e.g., 99% or more) (if the difference is less than the predetermined percentage), the data of both models may be considered to be substantially identical.

[0087] The identity of the learning parameters between the two models is determined by comparing the learning parameters excluding the number of times of learning. For example, if only the learning parameters that do not affect the model performance (such as the storage destination of the log data of the learning process) are different, the learning parameters of the two models may be considered to be identical.

[0088] The disclosure determination unit 105 performs a disclosure determination for the child model based on the disclosure criteria information output by the disclosure criteria information output unit 104. For example, if the determination result for the identity of the learning data and learning parameters of the parent model and child model is "TRUE" and the child model has been trained less than a predetermined number of times, the child model is deemed "private." This is because in this case, the child model has hardly progressed in learning from the parent model, and the two models can be considered to be substantially identical.

[0089] <Device Operation> Fig. 9 is a flowchart of processing in the information processing device 11 according to the third embodiment. Fig. 9(a) is a flowchart for learning and registering a model, and Fig. 9(b) is a flowchart for determining / updating the disclosure range of the model.

[0090] First, with reference to Fig. 9(a), a description will be given of the processing when a user operates the user terminal 14, registers data in the information processing device 11, and learns a model. Steps S901 to S904 are similar to steps S701 to S704 in the first embodiment (Fig. 7(a)), and therefore a description thereof will be omitted.

[0091] In S905, the disclosure determination unit 105 executes the model disclosure determination process shown in Fig. 9(b). That is, disclosure determination is performed at the same time as the model is registered. This makes it possible to prevent an inappropriate model (a child model that is substantially identical to a parent model) from being disclosed.

[0092] In S911, the model management unit 101 acquires traceability information from the traceability information management unit 103.

[0093] In S912, the model management unit 101 acquires model information of the model (child model) registered in S903 and its initial model (parent model) from the acquired traceability information. The acquired model information is passed to the disclosure standard information output unit 104.

[0094] In S913, the public reference information output unit 104 acquires the learning data ID and learning parameters from the model information of each of the parent model and the child model, and outputs the "determination result" of the identity of the learning data of both models and the "number of times the child model has learned" as public reference information.

[0095] In S914, the disclosure determination unit 105 determines whether to disclose the child model based on the disclosure criteria information. The detailed operation of the disclosure determination unit 105 is as described above.

[0096] In S915, the registration unit 106 updates the value of the disclosure range in the model information of the child model based on the result of the disclosure determination in S914.

[0097] In the above explanation, the disclosure determination of a registered model is performed at the same time as the new registration of the model, but this is not limited to this. For example, any model can be selected at any time using the administrator terminal 13, and the disclosure determination process shown in Figure 9(b) can be performed. In this case, if the disclosure determination result of the selected model is "private" or "limited disclosure," derived models of the same model may also be set to "private" or "limited disclosure" based on the traceability information.

[0098] As described above, according to the third embodiment, when a child model is newly registered in an information processing device, the degree of match between the child model and its parent model is confirmed, and the disclosure range of the child model is determined based on the degree of match. In particular, if the degree of match exceeds a predetermined threshold, the child model is set to "limited disclosure" or "private disclosure." This makes it possible to prevent a child model that is substantially identical to the parent model from being made public.

[0099] (Variation 3-1) In the above explanation, the public reference information output unit 104 outputs the "determination result" of the identity of the learning data of both models and the "number of times the child model has been learned" as public reference information, but other information may also be output as public reference information.

[0100] For example, object detection processing may be performed by both models on the evaluation image indicated by the evaluation data ID in the model information of the child model. Then, the IoU (Intersection over Union) between the detection result (BB) by the child model and the detection result (BB) by the parent model may be calculated, and the IoU value may be output as public reference information. IoU is an index that represents the percentage of overlap between the two detection results (BB), and is a value between 0 and 1, with a larger IoU indicating a greater percentage of overlap between the BBs. If there are multiple evaluation images, the average IoU value for each evaluation image may be output as public reference information.

[0101] In this case, if the IoU value (or average value) exceeds a predetermined value (for example, 0.99), the disclosure determination unit 105 evaluates that the similarity between the two models is high. In other words, if the detection performance of both models for a detection target appearing in the same image is comparable, the two models can be considered to be substantially identical.

[0102] (Variation 3-2) Furthermore, the public reference information output unit 104 may calculate the weight difference between the neural networks (NN) of the parent model and the child model (for each layer and kernel), and output the sum of the absolute values ​​of the differences as public reference information.

[0103] In this case, if the calculated sum is less than a predetermined value, the disclosure determination unit 105 evaluates that the similarity between the two models is high. In other words, since the difference in the weights of the NNs of the two models is small, the two models can be considered to be substantially the same.

[0104] (Variation 3-3) Furthermore, in model learning in the learning and evaluation unit 107, time-series data of loss during learning may be stored in the storage unit 12. Here, the loss is the difference between the predicted value by the NN and the true value of the learning data.

[0105] Then, the public reference information output unit 104 calculates the public reference information using the following formula (1) based on the learning rate of the learning parameters of the model information and the time-series data of the loss. Disclosure standard information = lr × Σ|∂Loss / ∂w| (1)

[0106] In formula (1), lr is the learning rate, Loss is the loss at a certain time, w is the weight of the model's NN, ∂ is the partial derivative, || is the absolute value, and Σ is the time accumulation. In other words, formula (1) is the value obtained by multiplying the time accumulation of the absolute value of the gradient of the loss by the learning rate. If this value is large, the NN weight will change significantly, and if it is small, the weight will change only slightly. In other words, the public reference information corresponds to the learning progress, which indicates how far the learning has progressed.

[0107] When the traceability information managed by the traceability information management unit 103 indicates that the model registered in S903 has a parent model, the disclosure determination unit 105 performs a disclosure determination for the model based on the above-mentioned disclosure standard information. If the calculated disclosure standard information is less than a predetermined value, it can be determined that the weight of the NN of the child model has hardly changed from the parent model, and the similarity between the two models can be determined to be high.

[0108] 10 is a flowchart for determining / updating the disclosure range of a model in Modification 3-3. The process of learning and registering a model is the same as in the third embodiment (FIG. 9(a)), so a description thereof will be omitted.

[0109] In S1011, the model management unit 101 acquires model information of a newly registered model from the storage unit 12 via the user terminal 14. The model information is passed to the disclosure criteria information output unit 104.

[0110] In S1012, the learning and evaluation unit 107 acquires time-series data of the loss during learning for the model registered in S903 from the storage unit 12. The time-series data of the loss is passed to the public reference information output unit 104.

[0111] In S1013, the public reference information output unit 104 calculates the public reference information using the formula (1).

[0112] In S1014 , the disclosure determination unit 105 acquires the traceability information managed by the traceability information management unit 103 .

[0113] In S1015, the disclosure determination unit 105 checks whether or not a parent model exists for the model registered in S903. If a parent model exists, the disclosure determination unit 105 determines whether or not the model registered in S903 is to be disclosed based on the disclosure standard information calculated in S1013.

[0114] In S1016, the registration unit 106 updates the value of the disclosure range in the model information of the model registered in S903 based on the result of the disclosure determination in S1015.

[0115] (Fourth embodiment) In the fourth embodiment, another information processing device that controls the disclosure range of a learning model will be described as an example. In particular, a graphical user interface (GUI) that manages traceability information provided by the display control unit 108 to the administrator terminal 13 and the user terminal 14 will be described. Note that it is assumed that the administrator terminal 13 and the user terminal 14 are configured by PCs, but this is not limited thereto. Also, the colors and line types on the GUI are merely examples and are not limited thereto.

[0116] <Traceability information management screen> 11 is a diagram showing an example of a traceability information management screen 1100. As described above, the management screen 1100 is provided to the administrator terminal 13 and / or the user terminal 14 by the display control unit 108. The management screen 1100 includes a traceability information screen 1110 and a detailed information screen 1120. For example, the detailed information screen 1120 displays detailed information about a GUI component selected by the user on the traceability information screen 1110.

[0117] The traceability information screen 1110 displays models 1101 to 1104 represented by rectangles, and the derivation relationships between the models are indicated by arrows in accordance with the traceability information (FIG. 5). Also displayed is a cursor 1105 for indicating a position on the management screen 1100. The cursor 1105 can be moved by the user operating, for example, a mouse (not shown). The traceability information screen 1110 also has buttons 1106 to 1108 for issuing instructions to execute various processes.

[0118] Models 1101 to 1104 each display a model ID and a disclosure range. Here, the disclosure range is displayed in text within a rectangle representing each model, but the disclosure range may also be displayed as a color within the rectangle. For example, a model with a disclosure range of "public (all)" may be displayed as a blue rectangle, a model with a disclosure range of "limited disclosure (to specific users)" may be displayed as a light blue rectangle, and model 1104 with a disclosure range of "private" may be displayed as a gray rectangle, etc.

[0119] The arrows connecting models indicate the derivation relationships between the learning models. The model connected to the root of the arrow is the initial model, and the model connected to the tip of the arrow is a derived model generated based on the initial model. That is, in Figure 11, model 1102 is a derived model of model 1101, model 1103 is a derived model of model 1102, and model 1104 is a derived model of model 1101. Model 1101 does not have a corresponding initial model, and represents an unlearned model (for example, a model in which the weights of the NN are set to random values).

[0120] 11 shows a state in which model 1101 is selected by cursor 1105. Here, the rectangle of model 1101 is displayed with a bold frame to clearly indicate that it is selected. Also in FIG. 11, the contents of the model information of model 1101 are displayed on detailed information screen 1120.

[0121] When a model is selected and the user presses (clicks) the mouse button on the learning button 1106, the screen transitions to a learning management screen (not shown), where learning processing can be started using the selected model 1101 as the initial model. Also, when a model is selected and the user clicks the evaluation button 1107, the screen transitions to an evaluation management screen (not shown), where evaluation processing can be performed on the selected model 1101 using the evaluation data described in the evaluation data ID. Furthermore, when a model is selected and the user clicks the disclosure determination button 1108, the screen transitions to a disclosure determination management screen (not shown), where disclosure determination processing can be performed on the selected model 1101 and its derived models.

[0122] As described above, according to the fourth embodiment, it is possible to provide (to the administrator terminal 13 and / or the user terminal 14) a GUI that enables the user to easily understand the publication status of each model and the derivation relationships between the models. Furthermore, via the GUI, the user can intuitively check detailed information about each model and perform various processes (learning, evaluation, and publication determination) for each model.

[0123] The disclosure of this specification includes the following information processing device, control method, and program. (Item 1) An information processing device that manages a plurality of learning models published on a model publishing platform, a management means for managing, for each of the plurality of learning models, first information on the learning history of the learning model and second information on the base model used to create the learning model; a determination means for determining a disclosure range of each of the plurality of learning models on the model disclosure platform based on both the first information and the second information; An information processing device comprising: (Item 2) The first information includes information about a dataset used in training a learning model; The determining means determines disclosure standard information indicating whether specific data is used in learning of each of the plurality of learning models based on the first information, and determines a disclosure range for each of the plurality of learning models based on the disclosure standard information corresponding to each of the plurality of learning models and the second information. 2. The information processing device according to item 1, (Item 3) The model publishing platform further includes a receiving unit for receiving the designation of the specific data from an administrator of the model publishing platform. 3. The information processing device according to item 2, (Item 4) the specified data is data with ethical and / or legal restrictions on use; The determining means determines, as the disclosure standard information, limited disclosure that narrows the disclosure range or non-disclosure that prohibits disclosure when the specific data is being used, and determines, as the disclosure standard information, full disclosure that does not impose any restrictions on the disclosure range when the specific data is not being used. 4. The information processing device according to item 2 or 3. (Item 5) The determining means determines the restricted release or the non-release for the learning model when the restricted release or the non-release has been determined for the base model identified by the second information used to create the learning model. 5. The information processing device according to item 4. (Item 6) Further, a verification means is provided for verifying whether the attention learning model has the model performance specified by the administrator of the model publishing platform; The determining means determines whether to make the attention learning model having the model performance publicly available by narrowing the public range or to make the model publicly available by prohibiting the model from being publicly available. 3. The information processing device according to item 2. (Item 7) The determination means determines, based on the first information, disclosure reference information indicating a difference between the learning history of the attention learning model and the learning history of a base model identified by the second information used to create the attention learning model, and determines the disclosure range of each of the plurality of learning models based on the disclosure reference information corresponding to each of the plurality of learning models and the second information. 2. The information processing device according to item 1, (Item 8) The first information includes a dataset and learning parameters used in learning the learning model, and a number of learning times; The determination means determines whether the attention learning model is to be made publicly available to a limited extent by narrowing the scope of disclosure or to be private by prohibiting disclosure, when a difference between the dataset and learning parameters used in learning the attention learning model and the dataset and learning parameters used in learning the base model is less than a predetermined ratio and the number of times the attention learning model has been learned is less than a predetermined number of times. 8. The information processing device according to item 7, (Item 9) The determining means determines the restricted release or the non-release of the learning model when the restricted release or the non-release has been determined for the base model used to create the learning model. 9. The information processing device according to item 8, (Item 10) the first information includes a detection result by a learning model for a given dataset; The determination means determines the IoU (Intersection over Union) between the detection result by the attention learning model and the detection result by the base model specified by the second information used to create the attention learning model as public reference information for the attention learning model, and determines the public range of each of the plurality of learning models based on the public reference information corresponding to each of the plurality of learning models and the second information. 2. The information processing device according to item 1, (Item 11) the first information includes weights of a neural network (NN) constituting each learning model; The determination means determines the sum of absolute values ​​of the differences between the weights of the NNs constituting the attention learning model and the weights of the NNs constituting the base model identified by the second information used to create the attention learning model as the public reference information of the attention learning model, and determines the public range of each of the plurality of learning models based on the public reference information corresponding to each of the plurality of learning models and the second information. 2. The information processing device according to item 1, (Item 12) the first information includes time-series data of a learning rate and a loss of a learning parameter during learning of each learning model; The determining means determines the learning progress calculated based on the learning rate of the noted learning model and the time-series data as disclosure standard information of the noted learning model, and determines the disclosure range of each of the plurality of learning models based on the disclosure standard information corresponding to each of the plurality of learning models and the second information. 2. The information processing device according to item 1, (Item 13) An information processing device that manages a plurality of learning models published on a model publishing platform, the plurality of learning models are object detection models for detecting objects from images, The information processing device includes: a registration means for registering keywords designated by an administrator of the model publishing platform in association with correct images for the keywords; a management means for managing, for each of the plurality of learning models, first information on the similarity between a generated image generated by inputting the keyword into the learning model and the correct image, and second information on a base model used to create the learning model; a determination means for determining a disclosure range of each of the plurality of learning models on the model disclosure platform based on both the first information and the second information; An information processing device comprising: (Item 14) a screen generation unit for generating a management screen to be displayed on an administrator terminal operated by an administrator of the model publishing platform; The management screen is configured to display, as a graphical user interface (GUI), derivation relationships between the plurality of learning models and publication states of each of the plurality of learning models based on the second information. 14. The information processing device according to any one of items 1 to 13, (Item 15) A control method for an information processing device that manages a plurality of learning models published on a model publishing platform, comprising: an acquisition step of acquiring, for each of the plurality of learning models, first information on the learning history of the learning model and second information on the base model used to create the learning model; a determination step of determining a disclosure range of each of the plurality of learning models on the model disclosure platform based on both the first information and the second information; A control method comprising: (Item 16) Item 16. A program for causing a computer to execute the control method according to Item 15.

[0124] (Other Examples) 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.

[0125] 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]

[0126] 10 control unit; 11 information processing device; 12 storage unit; 13 administrator terminal; 14 user terminal

Claims

1. An information processing device that manages a plurality of learning models published on a model publishing platform, a management means for managing, for each of the plurality of learning models, first information on the learning history of the learning model and second information on the base model used to create the learning model; a determination means for determining a disclosure range of each of the plurality of learning models on the model disclosure platform based on both the first information and the second information; An information processing device comprising:

2. The first information includes information about a dataset used in training a learning model; The determination means determines disclosure standard information indicating whether specific data is used in learning of each of the plurality of learning models based on the first information, and determines a disclosure range for each of the plurality of learning models based on the disclosure standard information corresponding to each of the plurality of learning models and the second information.

2. The information processing apparatus according to claim 1, wherein:

3. The model publishing platform further includes a receiving unit for receiving the designation of the specific data from an administrator of the model publishing platform.

3. The information processing apparatus according to claim 2, wherein:

4. the specific data is data whose use is restricted for ethical and / or legal reasons; The determining means determines, as the disclosure standard information, limited disclosure that narrows the disclosure range or non-disclosure that prohibits disclosure when the specific data is being used, and determines, as the disclosure standard information, full disclosure that does not impose any restrictions on the disclosure range when the specific data is not being used.

3. The information processing apparatus according to claim 2, wherein:

5. The determining means determines the restricted release or the non-release for the learning model when the restricted release or the non-release has been determined for the base model identified by the second information used to create the learning model.

5. The information processing apparatus according to claim 4,

6. Further, a verification means is provided for verifying whether the attention learning model has the model performance specified by the administrator of the model publishing platform; The determining means determines whether to make the attention learning model having the model performance publicly available by narrowing the public range or to make the model publicly available by prohibiting the model from being publicly available.

3. The information processing apparatus according to claim 2.

7. The determination means determines, based on the first information, disclosure reference information indicating a difference between the learning history of the attention learning model and the learning history of a base model identified by the second information used to create the attention learning model, and determines the disclosure range of each of the plurality of learning models based on the disclosure reference information corresponding to each of the plurality of learning models and the second information.

2. The information processing apparatus according to claim 1, wherein:

8. the first information includes a dataset and learning parameters used in learning a learning model, and a number of learning times; The determination means determines whether the attention learning model is to be made publicly available to a limited extent by narrowing the scope of disclosure or to be private by prohibiting disclosure, when a difference between the dataset and learning parameters used in learning the attention learning model and the dataset and learning parameters used in learning the base model is less than a predetermined ratio and the number of times the attention learning model has been learned is less than a predetermined number of times.

8. The information processing apparatus according to claim 7,

9. The determining means determines the restricted release or the non-release of the learning model when the restricted release or the non-release has been determined for the base model used to create the learning model.

9. The information processing apparatus according to claim 8,

10. the first information includes a detection result by a learning model for a given dataset; The determination means determines the IoU (Intersection over Union) between the detection result by the attention learning model and the detection result by the base model specified by the second information used to create the attention learning model as public reference information for the attention learning model, and determines the public range of each of the plurality of learning models based on the public reference information corresponding to each of the plurality of learning models and the second information.

2. The information processing apparatus according to claim 1, wherein:

11. the first information includes weights of neural networks (NNs) that configure each learning model; The determination means determines the sum of absolute values ​​of the differences between the weights of the NNs constituting the attention learning model and the weights of the NNs constituting the base model specified by the second information used to create the attention learning model as the public reference information of the attention learning model, and determines the public range of each of the plurality of learning models based on the public reference information corresponding to each of the plurality of learning models and the second information.

2. The information processing apparatus according to claim 1, wherein:

12. the first information includes time-series data of a learning rate and a loss of a learning parameter during learning of each learning model; The determining means determines the learning progress calculated based on the learning rate of the noted learning model and the time-series data as disclosure standard information of the noted learning model, and determines the disclosure range of each of the plurality of learning models based on the disclosure standard information corresponding to each of the plurality of learning models and the second information.

2. The information processing apparatus according to claim 1, wherein:

13. An information processing device that manages a plurality of learning models published on a model publishing platform, the plurality of learning models are object detection models for detecting objects from images, The information processing device includes: a registration means for registering keywords designated by an administrator of the model publishing platform in association with correct images for the keywords; a management means for managing, for each of the plurality of learning models, first information on the similarity between a generated image generated by inputting the keyword into the learning model and the correct image, and second information on a base model used to create the learning model; a determination means for determining a disclosure range of each of the plurality of learning models on the model disclosure platform based on both the first information and the second information; An information processing device comprising:

14. a screen generation unit for generating a management screen to be displayed on an administrator terminal operated by an administrator of the model publishing platform; The management screen is configured to display, as a graphical user interface (GUI), derivation relationships between the plurality of learning models and publication states of the plurality of learning models based on the second information.

14. The information processing apparatus according to claim 1, wherein the information processing apparatus is a computer.

15. A control method for an information processing device that manages a plurality of learning models published on a model publishing platform, comprising: an acquisition step of acquiring, for each of the plurality of learning models, first information on the learning history of the learning model and second information on the base model used to create the learning model; a determination step of determining a publication range of each of the plurality of learning models on the model publication platform based on both the first information and the second information; A control method comprising:

16. A program for causing a computer to execute the control method according to claim 15.

Images