Federated learning management device, federated learning management method, and program

Abstract

In the present invention, a contribution score of a data set is obtained for federated machine learning that is performed using data sets, and a remuneration amount for the data set is determined on the basis of at least the contribution score of the data set.

Description

Federated learning management device, federated learning management method, and program

[0001] The present invention relates to machine learning techniques, and more particularly to federated learning techniques.

[0002] Federated learning is a well-known technique for performing machine learning without aggregating datasets (learning data). In federated learning, multiple data providing devices perform machine learning (local learning) using datasets they hold to generate local models (worker models) and transmit the generated local models to the federated learning device. The federated learning device generates a global model (aggregated model) by aggregating the local models sent from the multiple data providing devices and transmits the generated global model to the multiple data providing devices. The multiple data providing devices that receive the global model update the global model using machine learning using datasets they hold to generate new local models and transmit the generated local models to the federated learning device. By repeating this process, machine learning can be performed without aggregating datasets.

[0003] Also, a method for calculating the contribution of a dataset used in machine learning is known (see, for example, Non-Patent Document 1).

[0004] Scott M. Lundberg, and Su-In Lee, "A Unified Approach to Interpreting Model Predictions," NIPS'17: Proceedings of the 31st International Conference on Neural Information Processing Systems, December 2017, Pages 4768-4777.

[0005] However, it is not known how to determine the reward amount for the dataset provided for federated learning.

[0006] The present invention provides a technique for determining the amount of reward for a dataset provided for federated learning.

[0007] For machine learning of federated learning performed using a dataset, the contribution of the dataset is obtained, and a reward amount for the dataset is determined based at least on the contribution of the dataset.

[0008] This allows us to determine the reward amount for the dataset provided for federated learning.

[0009] FIG. 1 is a block diagram illustrating the configuration of a federated learning system according to an embodiment. FIG. 2 is a block diagram illustrating the functional configuration of a federated learning management device according to an embodiment. FIG. 3 is a block diagram illustrating the functional configuration of a user device according to an embodiment. FIG. 4A is a block diagram illustrating the functional configuration of a data providing device according to an embodiment that also provides computational resources. FIG. 4B is a block diagram illustrating the functional configuration of a data providing device according to an embodiment that does not provide computational resources. FIG. 5 is a block diagram illustrating the functional configuration of a computational resource providing device according to an embodiment. FIG. 6 is a diagram illustrating the overall processing of federated learning according to an embodiment. FIG. 7 is a flow diagram illustrating registration processing that is a prerequisite for federated learning according to an embodiment. FIG. 8 is a flow diagram illustrating the processing of federated learning according to an embodiment. FIG. 9 is a flow diagram illustrating a method for determining computational resources for machine learning in federated learning according to an embodiment. FIG. 10 is a diagram illustrating a benchmark for federated learning. FIG. 11 is a flow diagram illustrating processing for determining computational resources to be used in machine learning according to an embodiment. FIG. 12 is a flow diagram illustrating processing for calculating and settling a reward for providing a dataset. FIG. 13 is a block diagram illustrating a hardware configuration of the embodiment.

[0010] [First Embodiment] An embodiment of the present invention will be described below with reference to the drawings. <Overall Configuration> As illustrated in FIG. 1, a federated learning system 1 of this embodiment includes a federated learning management device 11, a user device 12, data providing devices 13-1, ..., 13-N, and a computational resource providing device 14. Here, N is a positive integer, for example, N≧2. Each of the data providing devices 13-1, ..., 13-N will be referred to as data providing device 13-n (n∈{1, ..., N}). Of the data providing devices 13-1, ..., 13-N, those that provide computational resources will be referred to as data providing device 13-i, and those that do not provide computational resources will be referred to as data providing device 13-j (i, j∈{1, ..., N}). The data providing devices 13-1, ..., 13-N will be collectively referred to as data providing device 13. Hereinafter, the same applies to other notations for n, i, j∈{1, ..., N}. The federated learning management device 11, the user device 12, the data providing devices 13-1, ..., 13-N, and the computational resource providing device 14 can exchange information via a network. Examples of networks include an Innovative Optical and Wireless Network (IOWN), an All-Photonics Network (APN), and the Internet. When an IOWN or APN is used as the network, wavelengths are dynamically assigned to each communication path, and devices are directly connected via an optical communication network (path management control). Information is exchanged via the network while keeping the information confidential from third parties. Security technologies such as cryptography, secret sharing, or quantum communication are used for such information exchange (e.g., secure transport, secure federated learning, etc.). For simplicity of explanation, FIG. 1 illustrates one federated learning management device 11, one user device 12, and one computational resource providing device 14. However, there may be multiple federated learning management devices 11, one user device 12, and one computational resource providing device 14.

[0011] <Federated Learning Management Device 11> The federated learning management device 11 is a device for managing federated learning in this embodiment. As illustrated in Figure 2, the federated learning management device 11 in this embodiment includes a storage unit 110a, a memory 110b, a control unit 110c, a communication unit 111d, a model distribution management unit 110e, a model aggregation unit 110f, an authentication processing unit 111a, a consent management unit 111b, a data catalog management unit 111c, a CR (computing resource) determination unit 112, a learning control unit 114, and a settlement unit 115. The CR determination unit 112 includes a measurement unit 112a, a determination unit 112b, and a test CR unit 112c. The learning control unit 114 includes a measurement unit 114a, an evaluation unit 114b, and a CR management unit 114c. The settlement unit 115 has a contribution level acquisition unit 115a, a data provision fee calculation unit 115b, a CR usage fee calculation unit 115c, a remuneration amount determination unit 115d, and a settlement processing unit 115e. The federated learning management device 11 executes each process under the control of the control unit 110c. Information input to the federated learning management device 11 and information obtained in each process of the federated learning management device 11 are stored in memory 110b one by one, and are read out as needed and used in each process.

[0012] <User Device 12> The user device 12 is a device through which a user requests machine learning of a model and obtains a trained model. As illustrated in Fig. 3, the user device 12 of this embodiment has a storage unit 121, a collaboration processing unit 122, a communication unit 123, a UI (user interface) unit 124, a memory 125, and a control unit 126. The user device 12 executes each process under the control of the control unit 126. Information input to the user device 12 and information obtained in each process of the user device 12 are stored in the memory 125 one by one, read out as needed, and used in each process.

[0013] <Data Providing Device 13-i> The data providing device 13-i is a device that provides a dataset DS-i and computational resources for machine learning. The computational resources are, for example, a processor, memory, storage, software, a network, or any combination thereof. As illustrated in FIG. 4A, the data providing device 13-i has a storage unit 131-i, a collaboration processing unit 132-i, a communication unit 133-i, a CR (computational resource) unit 134-i, a memory 135-i, and a control unit 136-i. The storage unit 131-i stores a dataset DS-i provided by an entity (participant) using the data providing device 13-i. The data providing device 13-i executes each process under the control of the control unit 136-i. Information input to the data providing device 13-i and information obtained in each process of the data providing device 13-i are stored in the memory 135-i one by one, and are read out as needed and used in each process.

[0014] <Data providing device 13-j> The data providing device 13-j is a device that provides a dataset DS-j for machine learning but does not provide computational resources. As illustrated in FIG. 4B, the data providing device 13-j has a storage unit 131-j, a collaboration processing unit 132-j, a communication unit 133-j, a memory 135-j, and a control unit 136-j. The storage unit 131-j stores a dataset DS-j provided by an entity that uses the data providing device 13-j. The data providing device 13-j executes each process under the control of the control unit 136-j. Information input to the data providing device 13-j and information obtained in each process of the data providing device 13-j are stored in the memory 135-j one by one, and are read out as needed and used in each process.

[0015] <Computational Resource Providing Device 14> The computational resource providing device 14 is a device that provides computational resources for machine learning. As illustrated in FIG. 5, the computational resource providing device 14 of this embodiment includes a collaboration processing unit 142, a communication unit 143, CR (computational resource) units 144-1, ..., 144-M, a memory 145, and a control unit 146. Here, M is a positive integer, for example, M≧2. Each of the CR units 144-1, ..., 144-M is referred to as a CR unit 144-m (m∈{1, ..., M}), and the CR units 144-1, ..., 144-M are collectively referred to as a CR unit 144. Hereinafter, the same applies to other notations regarding m∈{1, ..., M}. The CR units 144-1, ..., 144-M of this embodiment have different performance (specifications). The computational resource providing device 14 executes each process under the control of the control unit 146. The information input to the computational resource providing device 14 and the information obtained in each process of the computational resource providing device 14 are stored in the memory 145 one by one, and are read out as needed and used in each process.

[0016] <Overall Federated Learning> The overall federated learning of this embodiment will be described using FIG. 6. First, a user uses a user device 12 (FIG. 1) to request the federated learning management device 11 to train a desired model. The federated learning management device 11 manages the federated learning of the requested model (global model) and transmits the finally obtained trained model to the user device 12. In this embodiment, the global model GM is updated by machine learning (local learning) using a dataset DS-n (n∈{1, ..., N}) specified by the user from among datasets DS-1, ..., DS-N, and each local model LM-n is generated. The user may specify a single dataset DS-n or multiple datasets DS-n. Machine learning using each dataset DS-n may be performed on the computational resources (CR) of a data providing device 13-i (i∈{1, ..., N}) (FIGS. 1 and 4A) or on the computational resources (CR) of a computational resource providing device 14 (FIGS. 1 and 5) (step S1-n). The generated local model or models LM-n are aggregated in the federated learning management device 11 (FIGS. 1 and 2) to generate a new global model GM. The generated global model GM is sent to each computational resource that performs machine learning in steps S1-n (step S2). The processes of steps S1-n and S2 are repeated until a termination condition is met, and a final trained model (global model GM) is obtained. Note that one model update cycle in which local models LM-n are generated by machine learning using each dataset DS-n, and the local models LM-n are aggregated to generate a global model GM is referred to as a round. Typically, in federated learning, multiple rounds of processing are performed to obtain a trained model.

[0017] <Registration Process> In the registration process, an entity that provides a dataset (training data) for machine learning is registered. The entity may be a corporation, a natural person, or another organization. In addition to the dataset, the entity may or may not provide computational resources for performing machine learning. This will be described in detail below with reference to FIG. 7.

[0018] The collaboration processing unit 132-n of the data providing device 13-n (FIGS. 4A and 4B) sends participation application information for registering an entity to the communication unit 133-n, and the communication unit 133-n transmits the participation application information to the federated learning management device 11 via the network. The participation application information is received by the communication unit 111d of the federated learning management device 11 (FIG. 2) and sent to the authentication processing unit 111a. The authentication processing unit 111a communicates with the collaboration processing unit 132-n of the data providing device 13-n (FIG. 4A or 4B) via the network, sets the ID and authentication information of the entity for which a data set is to be provided, and stores the registration information including these in the memory unit 110a (step S101).

[0019] Next, the collaboration processing unit 132-n of the data providing device 13-n (FIGS. 4A and 4B) sends consent management information to the communication unit 133-n, including information representing the available dataset DS-n, information representing the destination of the dataset DS-n, information representing whether or not there is an intention to provide computational resources, and information identifying the computational resources for which there is an intention to provide (only if there is an intention to provide computational resources). The information representing the available dataset DS-n may be the dataset DS-n itself, information representing a link destination of the dataset DS-n, information representing the attributes of the dataset DS-n, or information representing other datasets DS-n. Furthermore, the destination of the dataset DS-n may be a specific destination, all destinations, or all destinations excluding a specific destination. The destination may be classified by corporation, natural person, other organizational unit, field, industry, country, or other classification unit. The information specifying the intended computing resources to be provided includes information specifying the CR unit 134-i of the data providing device 13-i (FIG. 4A). The communication unit 133-n transmits the consent management information to the federated learning management device 11 via the network. The consent management information is received by the communication unit 111d of the federated learning management device 11 (FIG. 2) and sent to the consent management unit 111b. The consent management unit 111b adds the consent management information to the registration information of the corresponding entity and stores it in the memory unit 110a (step S102).

[0020] Furthermore, the consent management unit 111b refers to the registration information stored in the storage unit 110a and determines whether the entity providing the dataset DS-n also intends to provide computational resources (step S103). If the entity intends to provide computational resources, the consent management unit 111b adds to the registration information a statement indicating that the CR unit 134-i (second computational resource) of the entity's data providing device 13-i (FIG. 4A) will be set as a computational resource for machine learning, stores the information in the storage unit 110a (step S104), and terminates the registration process for the entity. On the other hand, if the entity does not intend to provide computational resources, the consent management unit 111b adds to the registration information a statement indicating that the entity will be loaned one of the CR units 144-m (first computational resource) of the computational resource providing device 14 (FIG. 5), stores the information in the storage unit 110a (step S105), and terminates the registration process for the entity. The CR unit 144-m (first computational resource) is a computational resource loaned to an entity that provides the dataset DS for machine learning. The CR unit 134-i (second computational resource) is a computational resource provided to an entity that provides the dataset DS-i for machine learning.

[0021] <User Registration Process> In the user registration process, a user requesting model learning is registered. The user may be a corporation, an individual, or another organization. The user enters user registration information required for user registration into the UI unit 124 of the user device 12 (Figure 3). The user registration information includes user attribute information used to determine whether the user is a recipient of a dataset. The user attribute information represents, for example, a corporation name, an individual name, an organization name, a field, an industry, a country, or other classifications. The user registration information is sent from the collaboration processing unit 122 to the communication unit 123. The communication unit 123 transmits the user registration information to the federated learning management device 11 via the network. The user registration information is received by the communication unit 111d of the federated learning management device 11 (Figure 2) and sent to the authentication processing unit 111a. The authentication processing unit 111a communicates with the collaboration processing unit 122 of the user device 12 (Figure 3) via the network, sets a user ID and authentication information, and stores the user registration information, including these and the user attribute information, in the memory unit 110a.

[0022] <Federated Learning> The federated learning process of this embodiment will be described with reference to FIG. 8. A user requesting model learning inputs learning request information, including a user ID, authentication information, information representing the model (global model) for which learning is requested, and information representing the deadline (delivery date), into the UI unit 124 of the user device 12 (FIG. 3). The information representing the model indicates, for example, the model type. Examples of models include logistic regression, random forest, gradient boosting, and deep neural network. The learning request information is sent to the collaboration processing unit 122 and further to the communication unit 123. The communication unit 123 transmits the learning request information to the federated learning management device 11 via the network. The learning request information is received by the communication unit 111d of the federated learning management device 11 (FIG. 2) and stored in the memory unit 110a. The authentication processing unit 111a performs user authentication processing using the user ID and authentication information included in the learning request information read from the memory unit 110a, as well as the user registration information stored in the memory unit 110a. If the user authentication process is successful, the learning request information is sent to the data catalog management unit 111c. The data catalog management unit 111c references the user attribute information included in the user registration information stored in the storage unit 110a and the consent management information included in the registration information, and sends a data catalog representing the datasets DS that can be provided to this user to the communication unit 110d. The communication unit 110d transmits the data catalog DS to the user device 12 via the network (step S111).

[0023] The data catalog is received by the communication unit 123 of the user device 12 (FIG. 3), sent to the collaboration processing unit 122, and then sent to the UI unit 124. The UI unit 124 outputs the data catalog to the user. The user refers to the data catalog, selects a dataset DS to use for learning, and inputs information for specifying the selected dataset DS into the UI unit 124. The information for specifying the dataset DS is sent to the collaboration processing unit 122 and transmitted from the communication unit 123 to the federated learning management device 11 via the network. The communication unit 111d of the federated learning management device 11 (FIG. 2) receives the information for specifying the dataset DS and sends it to the data catalog management unit 111c. As a result, the data catalog management unit 111c accepts the specification of the dataset DS (step S112).

[0024] Information representing the specified dataset DS is sent to the model distribution management unit 110e. The model distribution management unit 110e references the registration information stored in the storage unit 110a and sets computational resources for machine learning using the specified dataset DS. Here, for dataset DS-i, of the specified datasets DS, provided by an entity that also intends to provide computational resources (FIG. 7: step S104), the CR unit 134-i of that entity's data providing device 13-i (FIG. 4A) is set as the computational resource for machine learning. For dataset DS-j, of the specified datasets DS, provided by an entity that does not intend to provide computational resources (FIG. 7: step S105), one of the CR units 144-m of the computational resource providing device 14 (FIG. 5) is loaned to the entity and set as the computational resource for machine learning. The method for determining the CR unit 144-m (computational resource) to be loaned will be described later (step S113).

[0025] Information representing the global model GM included in the learning request information (e.g., model parameters) is sent to the model distribution management unit 110e. The model distribution management unit 110e sets the specified dataset and the initial state of the global model GM (e.g., initial model parameters) in the computational resources set in step S113. That is, the model distribution management unit 110e sets the specified dataset DS-i and the initial state of the global model GM in the CR unit 134-i of the data providing device 13-i (FIG. 4A), and sets the specified dataset DS-j and the initial state of the global model GM in the CR unit 144-m of the computational resource providing device 14 (FIG. 5). These settings are performed via a network, and the initial states of the dataset DS and the global model GM are transmitted via the network as necessary. The initial state of the global model GM may be, for example, predetermined or specified by input (step S114).

[0026] Under the management of the learning control unit 114 of the federated learning management device 11 (FIG. 2), each computational resource updates the global model GM through machine learning (local learning) using each dataset DS-n, thereby generating each local model LM-n. That is, the CR unit 134-i (FIG. 4A) or CR unit 144-m (FIG. 5), in which the dataset DS-n and the global model GM are set, updates the global model GM through machine learning using the dataset DS-n to generate a local model LM-n. The generated local model LM-n is sent to the communication unit 133-i (FIG. 4A) or communication unit 143 (FIG. 5) and transmitted to the federated learning management device 11 via the network. All generated local models LM-n are received by the communication unit 110d of the federated learning management device 11 (FIG. 2) and sent to the model aggregation unit 110f. The learning control unit 114 also evaluates the inadequacy of each computational resource for machine learning and switches out any inadequate computational resources. The management contents of each machine learning will be described later (step S115).

[0027] The model aggregating unit 110f generates a new global model GM by aggregating all the transmitted local models LM-n. There are no limitations on this aggregation process. For example, the model aggregating unit 110f generates a new global model GM in which the average value of each parameter of all the transmitted local models LM-n is used as the value of each parameter (step S116).

[0028] The model distribution management unit 110e determines whether the convergence condition is satisfied. There are no limitations on the convergence condition. Examples of the convergence condition include the improvement in the prediction accuracy of the local model LM-n or the global model GM before and after the round being below a standard, the change in the parameters of the local model LM-n or the global model GM before and after the round being below a standard, or the number of rounds reaching a standard number (step S117).

[0029] If it is determined that the convergence condition is not satisfied, a new global model GM is sent to the model delivery management unit 110e, and the process returns to step S115 (step S117).

[0030] On the other hand, if it is determined that the convergence condition is satisfied, the new global model GM is sent to the communication unit 110d. The communication unit 110d transmits this as the trained model M to the user device 12 via the network. The communication unit 123 of the user device 12 (FIG. 3) receives the trained model M and sends it to the collaboration processing unit 122. The collaboration processing unit 122 stores the trained model M in the storage unit 121 (step S118).

[0031] <Example of Method for Determining Computational Resources to be Loaned> FIG. 9 illustrates an example of the method for determining the computational resources to be loaned (step S113). When a CR unit 144-m (computational resource) is loaned to an entity, its performance must be determined. In other words, it is necessary to select a CR unit 144-m with performance suitable for the machine learning of the requested model from among the CR units 144-1, ..., 144-M. The performance of the computational resource refers to, for example, the number of processor cores, the number of threads, the clock frequency, the cache size, the memory capacity, and the network bandwidth. Here, the performance of the computational resource required for machine learning varies depending on the deadline and complexity of the requested model. In other words, the shorter the deadline (delivery date) of the trained model, the higher the performance of the computational resource required for machine learning. Furthermore, the more complex the model, the greater the amount of calculation required, and therefore the higher the performance of the computational resource required for machine learning. On the other hand, since the higher the performance of the computational resource, the higher the usage fee for the computational resource, it is not desirable to loan computational resources with higher performance than necessary. The model distribution management unit 110e takes these factors into consideration and determines the CR unit 144-m (computational resource) to be lent.

[0032] First, the test CR unit 112c of the CR determination unit 112 of the associative learning management device 11 (FIG. 2) initializes the test computational resources used to determine the computational resources to be lent. This initial setting may be determined in advance or by input (step S121).

[0033] The determination unit 112b extracts information representing the target model and information representing the deadline from the learning request information read from the storage unit 110a. Furthermore, the determination unit 112b determines the complexity of the target model. The determination of model complexity is performed, for example, based on correspondence information representing the relationship between models and model complexity. The correspondence information may be a table representing the relationship between models and model complexity, or may be a function. FIG. 10 illustrates an example table representing the relationship between models and model complexity. In the example of FIG. 10, model complexity is associated with each model, and machine learning time, communication time, total time per round, and target number of rounds until convergence according to each complexity are associated. These may be set based on past machine learning or simulation. For example, a model complexity of "simple" is associated with a model "logistic regression," which is associated with a machine learning time of "several seconds to several minutes," a communication time of "several seconds to 1 minute," a total time per round of "1 to 5 minutes," and a target number of rounds until convergence of "50 to 100 rounds." The model complexity "standard" is associated with the model "random forest, gradient boosting," which is associated with a machine learning time of "5 to 30 minutes," a communication time of "1 to 5 minutes," a total time per round of "10 to 35 minutes," and an estimated number of rounds until convergence of "100 to 200 rounds." The model complexity "complex" is associated with the model "deep neural net," which is associated with a machine learning time of "30 minutes to several hours," a communication time of "5 to 10 minutes," a total time per round of "35 minutes to several hours," and an estimated number of rounds until convergence of "200 to 500 rounds." In this example, if the target model is a "deep neural net," the model complexity is determined to be "complex" (step S122).

[0034] Next, the measurement unit 112a uses the test computational resources set by the test CR unit 112c to execute a process including a portion of the machine learning in federated learning on the target model, and measures the partial processing time. In other words, the partial processing time is the processing time required to execute a process including a portion of the machine learning in federated learning using the test computational resources. This machine learning may be performed, for example, using a dataset DS-n of the entity to which the computational resources are to be loaned, or may be performed using a dataset set for testing. The process including a portion of the machine learning may be any process. For example, the process including a portion of the machine learning may be a single-round process to generate a local model and a global model, two or more rounds to generate a local model and a global model, a process to generate a local model using one round of machine learning, or a process to generate a local model using two or more rounds of machine learning. The measured partial processing time is sent to the determination unit 112b (step S123).

[0035] The determination unit 112b estimates whether a trained model M can be obtained within the deadline using the deadline and complexity of the model obtained in step S122 and the partial processing time obtained in step S123. For example, the determination unit 112b estimates the number of rounds required to perform all federated learning of the target model based on the complexity of the target model (e.g., FIG. 10), and then estimates the total time required to perform all federated learning of the target model based on the estimated number of rounds and the partial processing time obtained in step S123, and estimates whether this time is within the deadline (step S124). Here, if it is estimated that a trained model M cannot be obtained within the deadline, the test CR unit 112c increases the performance of the test computing resources so as to shorten the partial processing time. In other words, the test CR unit 112c increases the performance of the test computing resources so as to increase the execution speed of machine learning (step S125). Then, the process returns to step S123.

[0036] On the other hand, if it is estimated in step S124 that a trained model M can be obtained within the deadline, the test computational resources can be said to have performance corresponding to the partial processing time estimated to complete the federated learning within the deadline. In this case, the determination unit 112b further determines whether the partial processing time is within the range expected for the model complexity based on the partial processing time and the complexity of the target model (step S126). The determination unit 112b determines the performance of the computational resources to be loaned based on this determination result. In other words, even if the partial processing time is sufficient to meet the deadline, if it exceeds the range expected for the model complexity, the overall performance of the federated learning will be reduced. Furthermore, if the partial processing time is below the range expected for the model complexity, the performance of the computational resources is excessive. Because the global model GM is updated after multiple local models LM are obtained, even if only some of the local models LM are generated in a short time, the overall performance will not improve. On the other hand, the higher the performance of the computational resource, the higher the usage fee. Therefore, the determination unit 112b determines the performance of the computational resources to be loaned so that the partial processing time is within the range expected for the complexity of the target model. In other words, if the partial processing time exceeds the range expected based on the model's complexity, the determination unit 112b selects a CR unit 144-m (computational resource) with higher performance than the test computational resource, and terminates processing (step S127). If the partial processing time falls below the range expected based on the model's complexity, the determination unit 112b selects a CR unit 144-m (computational resource) with lower performance than the test computational resource, and terminates processing. Even in this case, the CR unit 144-m must have sufficient performance to meet the deadline (step S128). On the other hand, if the partial processing time falls within the range expected based on the model's complexity, the determination unit 112b selects a CR unit 144-m (computational resource) with the same or similar performance as the test computational resource, and terminates processing (step S129).

[0037] In this manner, the determination unit 112b of this embodiment determines the machine learning computational resources to be lent based on the partial processing times. Preferably, the determination unit 112b of this embodiment determines the machine learning computational resources to be lent based on the performance of the test computational resources corresponding to the partial processing times estimated to complete the federated learning within the deadline (step S124). More preferably, the determination unit 112b of this embodiment determines the machine learning computational resources to be lent based on the performance of the test computational resources corresponding to the partial processing times within the range expected from the complexity of the model that is the target of the federated learning (steps S126 to S129).

[0038] <Example of Management Contents of Each Machine Learning> An example of the management contents of each machine learning (step S115) is shown using Figure 11. The CR management unit 114c of the learning control unit 114 of the federated learning management device 11 (Figure 2) determines whether any of the CR units 134-i (computational resources) (Figure 4A) have been configured for machine learning using dataset DS-n (step S131). Here, if any of the CR units 134-i have not been configured for machine learning using dataset DS-n, the CR management unit 114c causes the loaned CR unit 144-m (computational resources) (Figure 5) to perform machine learning using the dataset DS-n (step S132).

[0039] On the other hand, if any CR unit 134-i (computational resource) (FIG. 4A) is configured for machine learning using dataset DS-n, the measurement unit 114a (second measurement unit) of the learning control unit 114 measures the partial processing time (second partial processing time) of the provided CR unit 134-i (computational resource). That is, the measurement unit 114a executes a process including a part of the machine learning using the provided CR unit 134-i (computational resource) and measures the partial processing time, which is the processing time. The process including a part of the machine learning may be the same as or different from that in step S123. For example, the process including a part of the machine learning may be a single-round process for generating a local model and a global model, or two or more rounds for generating a local model and a global model, or a process for generating a local model through one round of machine learning, or a process for generating a local model through two or more rounds of machine learning. Furthermore, this measurement may be performed for each round or for multiple rounds. The measured partial processing time is sent to the evaluation unit 114b (step S133).

[0040] The evaluation unit 114b uses the sent partial processing time (second partial processing time) to evaluate whether the provided CR unit 134-i (computational resource) is suitable for machine learning. Various methods can be considered for this evaluation. Specific examples are shown below.

[0041] Specific example 1: As in step S124, the evaluation unit 114b uses the deadline and complexity of the model obtained in step S122 and the partial processing time obtained in step S133 to estimate whether a trained model M can be obtained in time for the deadline, and if a trained model M can be obtained in time for the deadline, evaluates the CR unit 134-i as being suitable for machine learning, and if not, evaluates it as being unsuitable for machine learning.

[0042] Specific Example 2: The evaluation unit 114b uses the partial processing times obtained in step S133 to determine whether the machine learning using the provided CR unit 134-i is a bottleneck in the overall federated learning. If it is not a bottleneck, the evaluation unit 114b evaluates that the CR unit 134-i is suitable for machine learning. If not, the evaluation unit 114b evaluates that the CR unit 134-i is not suitable for machine learning. Note that the fact that the machine learning using the provided CR unit 134-i is a bottleneck in the overall federated learning means that the processing time of the machine learning using the CR unit 134-i is particularly long among the machine learning (local learning) using each computational resource. For example, the evaluation unit 114b may determine that the machine learning using the CR unit 134-i is a bottleneck if the length of the processing time of the machine learning using the CR unit 134-i is within a predetermined length (e.g., the longest processing time) among the machine learning (local learning) of the federated learning. Alternatively, for example, the evaluation unit 114b may determine that machine learning using the CR unit 134-i is a bottleneck if the length of the processing time for machine learning using the CR unit 134-i is longer than a predetermined time period than the average length of the processing time for machine learning using federated learning (local learning).

[0043] Specific example 3: The evaluation unit 114b uses the partial processing time obtained in step S133 to compare the processing time of machine learning using the CR unit 134-i with that of a past round (e.g., the previous round), and if the processing time is longer than a reference time compared to the past round, it determines that the CR unit 134-i is not suitable for machine learning; otherwise, it evaluates that it is suitable for machine learning.

[0044] That is, the evaluation unit 114b may evaluate whether the computational resources are suitable for machine learning based on an absolute criterion using partial processing time (e.g., specific example 1), or may evaluate whether the computational resources are suitable for machine learning based on a relative criterion compared with the processing time of other machine learning (e.g., specific example 2), or may evaluate whether the computational resources are suitable for machine learning based on a relative criterion compared with past processing time (e.g., specific example 3) (step S134).

[0045] Here, if the CR unit 134-i is evaluated as being suitable for machine learning, the CR management unit 114c of the learning control unit 114 (Figure 2) causes the CR unit 134-i to perform machine learning using the dataset DS-i (step S135).

[0046] On the other hand, if the provided CR unit 134-i (computational resource) is evaluated as not suitable for machine learning, the CR management unit 114c of the learning control unit 114 (FIG. 2) lends the CR unit 144-m (FIG. 5) to the entity that provided the CR unit 134-i, and switches the CR unit 134-i to the CR unit 144-m. Note that the method for determining the CR unit 144-m is, for example, the same as the computational resource determination method (step S113) described above. Thereafter, the CR management unit 114c executes machine learning using the dataset DS-i with the switched CR unit 144-m. In other words, if the learning control unit 114 (FIG. 2) determines that the provided computational resource (CR unit 134-i) (second computational resource) is not suitable for machine learning, it performs processing to execute machine learning using the loaned computational resource (CR unit 144-m) (first computational resource). This prevents machine learning from being hindered due to insufficient provided computational resources, enabling smooth federated learning. Furthermore, even if the learning control unit 114 (FIG. 2) determines that the provided computational resources (CR unit 134-i) are suitable for machine learning and performs processing to execute machine learning using the provided computational resources (CR unit 134-i), if it determines that the provided computational resources (CR unit 134-i) (second computational resources) are no longer suitable for machine learning, it switches (fails over) the provided computational resources (CR unit 134-i) (second computational resources) to the loaned computational resources for machine learning (CR unit 144-m) (first computational resources), and performs processing to execute machine learning using the loaned computational resources for machine learning (CR unit 144-m) (first computational resources). This prevents the performance of computational resources that were once accepted as suitable for machine learning from deteriorating during machine learning, causing problems with machine learning. In subsequent rounds, the CR unit 134-i may be considered not to be configured for machine learning of the data set DS-i, or the CR unit 134-i may be considered to be configured for machine learning of the data set DS-i (step S136).

[0047] <Payment Processing> Next, the payment processing of this embodiment will be described. In this embodiment, for machine learning through federated learning performed using a provided dataset DS, the contribution of the dataset DS is obtained, and a remuneration amount for providing the dataset DS is determined based at least on the contribution of the dataset DS. Furthermore, when processing for lending computational resources is performed and machine learning is performed using the loaned computational resources, a remuneration amount is further determined based on the usage fee for the loaned computational resources.

[0048] 12 illustrates a method for determining the amount of reward for an entity that provided dataset DS-n. The contribution level acquisition unit 115a of the settlement unit 115 of the federated learning management device 11 (FIG. 2) extracts information representing dataset DS-n from the registered information in the storage unit 110a and obtains the contribution level of dataset DS-n for the machine learning of federated learning performed using dataset DS-n. The contribution level of dataset DS-n may be evaluated based on the quantity of dataset DS-n, the quality, or a combination of these. Specific examples are provided below.

[0049] Evaluation based on the amount of dataset DS-n is an effective evaluation method when multiple entities provide datasets DS with the same items (attributes). However, evaluation based on the amount of dataset DS-n may also be performed when multiple entities provide datasets DS with different items. The contribution level acquisition unit 115a may evaluate the greater the contribution of dataset DS-n the greater the data amount of dataset DS-n, or may evaluate the greater the contribution of dataset DS-n the greater the number of data items included in dataset DS-n, or may evaluate the greater the contribution of dataset DS-n the greater the ratio or difference between the data amount or number of data of dataset DS-n relative to the entire dataset DS used in federated learning.

[0050] The provided dataset DS-n may be missing, contain errors, or be out of date. In such cases where the quality of the dataset DS-n is low, the contribution degree acquiring unit 115a may lower the contribution degree of the dataset DS-n. For example, when the quality of the provided dataset DS-n does not meet a standard, the contribution degree acquiring unit 115a may determine the contribution degree of the dataset DS-n by multiplying the contribution degree based on the quantity of the dataset DS-n by a positive coefficient less than 1. Such a coefficient may be a constant or a variable corresponding to the quality of the dataset DS-n (a variable that becomes smaller the lower the quality). Alternatively, for example, the contribution degree acquiring unit 115a may evaluate the contribution degree of the dataset DS-n as low when the quality of the provided dataset DS-n does not meet a standard, regardless of the quantity of the dataset DS-n. For example, the contribution degree acquiring unit 115a may set the contribution degree according to the quality of the dataset DS-n (a contribution degree that becomes smaller the lower the quality) regardless of the quantity of the dataset DS-n. Quality criteria such as defects, errors, freshness, etc. may be predetermined or input.

[0051] Conversely, there are cases where the dataset DS-n has a small variation (distribution), covers a wide variety of data, or includes rare data. In such cases where the quality of the dataset DS-n is high, the contribution degree acquiring unit 115a may increase the contribution degree of the dataset DS-n. For example, when the quality of the provided dataset DS-n exceeds a standard, the contribution degree acquiring unit 115a may determine the contribution degree of the dataset DS-n by multiplying the contribution degree based on the quantity of the dataset DS-n by a positive coefficient greater than 1. Such a coefficient may be a constant, or a variable corresponding to the quality of the data in the dataset DS-n (a variable that increases as the data quality increases). Alternatively, for example, the contribution degree acquiring unit 115a may highly evaluate the contribution degree of the dataset DS-n when the quality of the provided dataset DS-n exceeds a standard, regardless of the quantity of the dataset DS-n. For example, the contribution degree acquiring unit 115a may set the contribution degree according to the quality of the dataset DS-n (a contribution degree that increases as the quality increases), regardless of the quantity of the dataset DS-n. The quality criteria, such as variability (distribution), diversity, and rarity, may be predetermined or may be input.

[0052] The contribution of dataset DS-n may be evaluated based on the extent to which machine learning using dataset DS-n contributes to improving the prediction accuracy of global model GM. This contribution evaluation is an effective evaluation method not only when multiple entities provide datasets DS of the same item, but also when they provide datasets DS of different items. For example, consider a case in which multiple datasets DS used in federated learning represent features of different attributes, and a machine learning model is developed that uses these multiple features as inputs and outputs prediction results. In this case, the contribution acquisition unit 115a, for example, determines the weight of each feature according to its contribution by evaluating the contribution of each feature to the prediction result using the trained model M, and then divides the weight of the feature by the sum of the weights of all the feature values ​​to determine the contribution ratio of each feature value. The contribution ratio of each feature value is then used as the contribution ratio of the dataset DS-n corresponding to each feature value. To determine the weight of each feature, for example, Shapley values ​​(see, for example, Non-Patent Document 1), permutation importance, partial dependence plots, localized predictive interpretation (LIME), etc. can be used. Alternatively, a function value of the Shapley values ​​can be used. Furthermore, when multiple datasets DS represent features of the same attribute, the contribution of each dataset DS corresponding to the feature can be determined based on, for example, the contribution ratio of the feature obtained as described above and the ratio of the data volume of the multiple datasets DS corresponding to the feature. For example, when datasets DS-1 and DS-2 represent features of the same attribute and the ratio of the data volume between datasets DS-1 and DS-2 is 3:7, the contribution obtaining unit 115a can determine 3 / 10 of the contribution ratio of the feature obtained as described above as the contribution of dataset DS-1 and 7 / 10 as the contribution of dataset DS-2. Alternatively, the contribution of each data set DS-n corresponding to the feature may be determined based on the quality of the data, rather than just the proportion of the data amount.For example, by using the partial dependence plot technique, it is possible to fix variables other than a specific feature and check the degree of impact on the prediction results when the feature is varied, thereby making it possible to distinguish between ranges of feature quantities with the same attribute that have a significant impact on the prediction and ranges that do not. Therefore, even among feature quantities with the same attribute, the contribution of a dataset DS in a range that has a significant impact on the prediction can be made greater than the contribution of datasets DS in other ranges of the feature. In this way, by differentiating the contribution of datasets DS in a range that has a significant impact on the prediction results of the trained model M from the contribution of datasets DS in other ranges, the contribution of each dataset DS-n corresponding to the feature can be determined based not only on the proportion of data volume but also on the quality of the data.

[0053] The contribution degree of each data set DS-n obtained by the contribution degree obtaining unit 115a is sent to the data provision fee calculating unit 115b (step S141).

[0054] The data provision fee calculation unit 115b calculates the data provision fee for each entity's dataset DS-n based on the degree of contribution of each dataset DS-n. For example, the data provision fee calculation unit 115b sets the function value of the degree of contribution of each dataset DS-n (for example, a non-decreasing function value or a monotonically increasing function value) as the data provision fee for that dataset DS-n. For example, the data provision fee calculation unit 115b calculates a higher data provision fee for a dataset DS-n with a greater degree of contribution. The data provision fee for each entity's dataset DS-n is sent to the remuneration determination unit 115d (step S142).

[0055] The CR usage fee calculation unit 115c (FIG. 2) determines whether the CR unit 144-m (computational resources) was loaned to the entity that provided the data set DS-n, and calculates the usage fee for the CR unit 144-m (computational resources). The usage fee for the CR unit 144-m may be determined based on the performance of the loaned CR unit 144-m (FIG. 9: S121 to S129). For example, the higher the performance of the loaned CR unit 144-m, the higher the usage fee for the CR unit 144-m. In other words, when machine learning is performed using the CR unit 144-m (computational resources for machine learning), the CR usage fee calculation unit 115c may determine the usage fee for the CR unit 144-m based at least on the type of the CR unit 144-m. Furthermore, the CR usage fee calculation unit may calculate the usage fee for the CR unit 144-m based on the usage time of the loaned CR unit 144-m. For example, the longer the learning time by the loaned CR unit 144-m, the higher the usage fee for the CR unit 144-m may be. Also, if some rounds of machine learning using the dataset DS-i are executed by the provided CR unit 134-i (FIG. 4A) and other rounds are executed by the loaned CR unit 144-m (FIG. 5) (if it is determined that the provided computational resources are not suitable for machine learning) (FIG. 11: steps S134 to S136), the usage fee for the CR unit 144-m may be determined based on these learning times. In other words, the CR usage fee calculation unit 115c calculates the time T during which machine learning is executed using the provided CR unit 134-i (computational resources). ｉ and the time T during which machine learning is performed using the loaned CR unit 144-m (computational resource). ｍ For example, the CR fee calculation unit 115c may determine the usage fee of the CR unit 144-m based on the ratio T ｍ / (T ｉ +T ｍ The larger the difference T ｍ -T ｉThe larger the value of the CR unit 144-m, the higher the usage fee for the CR unit 144-m. The usage fee for the CR unit 144-m of an entity to which the CR unit 144-m has not been loaned may be zero or a fixed amount (minimum usage fee). The usage fee for the CR unit 144-m of each entity is sent to the remuneration determination unit 115d and the settlement processing unit 115e (step S143).

[0056] The remuneration determination unit 115d calculates the remuneration for each entity's provision of dataset DS-n based on the data provision fee for each entity's dataset DS-n and the usage fee for each entity's CR unit 144-m. For example, the remuneration determination unit 115d calculates the amount obtained by subtracting the usage fee for each entity's CR unit 144-m from the data provision fee for dataset DS-n of each entity as the remuneration for each entity's provision of dataset DS-n. The remuneration for the provision of dataset DS-n of an entity whose usage fee for CR unit 144-m is zero is, for example, the same as the data provision fee for dataset DS-n of that entity. In other words, the remuneration determination unit 115d determines the remuneration for the provision of dataset DS-n based at least on the degree of contribution of dataset DS-n. Furthermore, if the CR unit 144-m has been loaned, the remuneration determination unit 115d may determine the remuneration for providing the dataset DS-n based on the contribution of the dataset DS-n as well as the usage fee for the loaned CR unit 144-m (computational resources). In this case, the remuneration determination unit 115d may determine the remuneration for providing the dataset DS-n based on the usage fee for the CR unit 144-m determined based at least on the type of CR unit 144-m. For example, if the provided CR unit 134-i (computational resources) is determined to be unsuitable for machine learning and the computational resources are switched, the remuneration determination unit 115d may determine the remuneration for providing the dataset DS-n based on the contribution of the dataset DS-n as well as the usage fee for the switched CR unit 144-m (computational resources). In this case, the remuneration determination unit 115d may determine the remuneration for providing the data set DS-n based on the time for which machine learning was performed using the provided CR unit 134-i (computational resources) and the time for which machine learning was performed using the loaned CR unit 144-m (computational resources). The remuneration for providing each entity's data set DS-n is sent to the settlement processing unit 115e (step S144).

[0057] The settlement processing unit 115e performs settlement processing to give each entity a reward for providing the data set DS-n (step S144), settlement processing to pay the usage fee for the CR unit 144-m to the provider of the computational resource providing device 14, and settlement processing to collect the usage fee for the federated learning management device 11 from the user (step S145).

[0058] [Features of the Present Embodiment] In the present embodiment, for federated learning machine learning performed using a provided dataset, the contribution of the dataset is obtained, and a reward for providing the dataset is determined based at least on the contribution of the dataset. This makes it possible to automatically determine an appropriate reward for an entity that provides a dataset for federated learning machine learning. Furthermore, by determining the reward based on the contribution, rewards can be appropriately allocated when federated learning is performed using datasets provided by multiple entities. Furthermore, each entity can receive a reward based on the contribution of the dataset it provided, which motivates them to provide many or excellent datasets and improves the quality of the federated learning overall. If the contribution of a dataset is evaluated not only by quantity but also by quality, the quality of the federated learning can be further improved. Furthermore, if entities that can provide high-quality datasets can receive high rewards, the asset value of data based on quality can be objectively determined, which is expected to lead to new business developments. Furthermore, the value of the same dataset may vary depending on the model trained using it. Even in such cases, by evaluating the contribution of the dataset to a trained model, the asset value of the dataset according to the type of model can be appropriately evaluated and appropriate rewards can be provided.

[0059] Preferably, in this embodiment, when a process for lending computational resources to an entity that provided a dataset is performed and a process for executing machine learning using the loaned computational resources is performed, a remuneration amount for the entity that provided the dataset is determined based on the usage fee for the loaned computational resources. This makes it possible to automatically determine an appropriate remuneration amount that takes into account the usage fee for the loaned computational resources for an entity that can provide a dataset but cannot provide appropriate computational resources.

[0060] Preferably, in this embodiment, if it is determined that the provided computational resources are not suitable for machine learning, a process is performed to execute machine learning using the loaned computational resources, and a remuneration amount is determined based on the usage fee for the loaned computational resources. This prevents machine learning from being hindered due to insufficient provided computational resources, enables smooth federated learning, and automatically determines an appropriate remuneration amount that takes into account the usage fee for the loaned computational resources.

[0061] Preferably, in this embodiment, a process for executing machine learning using provided computational resources is performed, and if it is determined that the provided computational resources are no longer suitable for machine learning, the provided computational resources are switched to loaned computational resources, and a process for executing machine learning using the loaned computational resources is performed. Furthermore, a remuneration amount is determined based on the usage fee for the loaned computational resources. This prevents the performance of computational resources once accepted as suitable for machine learning from deteriorating during machine learning, which would hinder the machine learning, and also automatically determines an appropriate remuneration amount that takes into account the usage fee for the loaned computational resources. In this case, the remuneration amount may be determined based on the time spent executing machine learning using the loaned computational resources, or the time spent executing machine learning using the provided computational resources and the time spent executing machine learning using the loaned computational resources. This allows a more appropriate remuneration amount to be determined.

[0062] Preferably, in this embodiment, a partial processing time, which is the processing time required to execute a process including a part of machine learning using the provided computational resources, is measured, and whether the provided computational resources are suitable for machine learning is determined based on the partial processing time. This prevents machine learning from being hindered due to insufficient provided computational resources, and enables smooth federated learning to be implemented.

[0063] In this embodiment, the partial processing time, which is the processing time required when a process including a part of machine learning in federated learning is executed using test computational resources, is measured, and the computational resources for machine learning to be lent are determined based on the partial processing time. This makes it possible to lend appropriate computational resources according to the target model.

[0064] In this embodiment, the computational resources for machine learning are preferably determined based on the performance of the computational resources for testing that correspond to a partial processing time that is estimated to complete the federated learning within the deadline. This allows appropriate computational resources to be lent when a deadline is specified.

[0065] Preferably, in this embodiment, when machine learning is performed using the loaned computational resources, the usage fee for the computational resources is determined based on at least the type (performance) of the loaned computational resources, thereby making it possible to appropriately determine the usage fee for computational resources having the required performance.

[0066] [Variation of the First Embodiment] In the first embodiment, when machine learning is performed using loaned computational resources, the remuneration amount for the entity that provided the dataset is determined based on the usage fee for the loaned computational resources. However, even when machine learning is performed using loaned computational resources, the remuneration amount for the entity that provided the dataset may be determined without taking into account the usage fee for the loaned computational resources.

[0067] In the first embodiment, the computational resources of the computational resource providing device 14 are lent, but instead, the computational resources of any of the data providing devices 13 may be lent. In this case, the fee for providing the computational resources of the data providing device 13 may be further taken into consideration when determining the amount of remuneration to the entity that lent the computational resources.

[0068] In step S113 (FIG. 8) of the first embodiment, if an entity providing a dataset intends to provide computational resources, the CR unit 134-i of that entity's data providing device 13-i (FIG. 4A) is set as a computational resource for machine learning. However, if there is an intention to provide computational resources, the computational resources may be evaluated for suitability for machine learning, and the provision of computational resources that do not meet the criteria may be refused. An entity that refuses to provide computational resources may be loaned one of the CR units 144-m of the computational resource providing device 14 (FIG. 5), and the loaned CR unit 144-m may be set as a computational resource for machine learning. If the target model and deadline are known, the suitability of the provided computational resources for machine learning may be evaluated, for example, using the method illustrated in step S134.

[0069] In the method for determining the computational resources to be lent illustrated in FIG. 9, whether or not the trained model M can be obtained within the deadline is estimated (step S124), and if it is estimated that the trained model M cannot be obtained within the deadline, the performance of the test computational resources is increased (step S125). However, if no deadline is set, the processes of steps S124 and S125 may be omitted. Conversely, if a deadline is set, the processes of steps S126, S127, and S128 may be omitted, and if it is estimated in step S124 that the trained model M can be obtained within the deadline, a CR unit 144-m (computational resource) with the same or similar performance as the test computational resource may be selected. Furthermore, in the first embodiment, the test CR unit 112c of the CR determination unit 112 was used to determine the computational resources to be lent (FIG. 9). However, an external computational resource may be used instead of the test CR unit 112c of the CR determination unit 112.

[0070] The federated learning management device 11 may have a rough estimation unit that calculates a remuneration amount for providing the aforementioned dataset DS (steps S141 to S144) before starting federated learning of the target model or during the federated learning, and outputs the calculated remuneration amount as a rough estimate. Furthermore, if, during the federated learning, it is determined that the computational resources provided by the data providing device 13 are not suitable for machine learning and the machine learning is continued by switching to computational resources from the computational resource providing device 14 (step S136), the rough estimation unit may calculate a remuneration amount taking into account the switch in computational resources (steps S141 to S144) and output the newly calculated remuneration amount as a rough estimate. By transmitting such a rough estimate to the user device 12 via the network, the user can know the approximate cost required for model learning before the federated learning is completed.

[0071] [Hardware Configuration] The functions performed by the components described herein may be implemented in circuitry or processing circuitry, including general-purpose processors, application-specific processors, integrated circuits, ASICs (Application Specific Integrated Circuits), CPUs (Central Processing Units), conventional circuits, and / or combinations thereof, programmed to perform the described functions. A processor includes transistors and other circuits and is considered to be circuitry or processing circuitry. A processor may also be a programmed processor that executes a program stored in a memory.

[0072] In this specification, a circuitry, unit, or means is hardware that is programmed to realize or performs the described functions, which may be any hardware disclosed herein or any hardware known to be programmed to realize or perform the described functions.

[0073] If the hardware is a processor considered to be a type of circuitry, the circuitry, means, or unit is a combination of the hardware and software used to configure the hardware and / or processor.

[0074] For example, each device in each embodiment is configured by a general-purpose or dedicated computer having a processor (hardware processor) such as a CPU (Central Processing Unit) and memories such as RAM (Random-Access Memory) and ROM (Read-Only Memory) executing a predetermined program. That is, each device in each embodiment has, for example, a processing circuit configured to implement each unit. This computer may have one processor and memory, or multiple processors and memories. This program may be installed on the computer or may be pre-recorded in a ROM or the like. Furthermore, some or all of the processing units may be configured using electronic circuits that independently realize processing functions, rather than electronic circuits that realize functional configurations by loading programs, such as a CPU. Furthermore, the electronic circuits that make up a single device may include multiple CPUs.

[0075] FIG. 13 is a block diagram illustrating the hardware configuration of a computer according to this embodiment. As illustrated in FIG. 13, the computer of this example includes a central processing unit (CPU) 10a, an input unit 10b, an output unit 10c, a random access memory (RAM) 10d, a read-only memory (ROM) 10e, an auxiliary storage device 10f, a communication unit 10h, and a bus 10g. The CPU 10a of this example includes a control unit 10aa, a calculation unit 10ab, and a register 10ac, and executes various calculation processes according to various programs loaded into the register 10ac. The input unit 10b is an input terminal, keyboard, mouse, touch panel, or the like, through which data is input. The output unit 10c is an output terminal, display, or the like, through which data is output. The communication unit 10h is a LAN card or the like, controlled by the CPU 10a that has loaded a predetermined program. The RAM 10d is a static random access memory (SRAM), dynamic random access memory (DRAM), or the like, and has a program area 10da where predetermined programs are stored and a data area 10db where various data are stored. The auxiliary storage device 10f is a hard disk, magneto-optical disc (MO), semiconductor memory, or the like, and has a program area 10fa where predetermined programs are stored and a data area 10fb where various data are stored. The bus 10g connects the CPU 10a, input unit 10b, output unit 10c, RAM 10d, ROM 10e, communication unit 10h, and auxiliary storage device 10f so that information can be exchanged. The CPU 10a writes the program stored in the program area 10fa of the auxiliary storage device 10f to the program area 10da of RAM 10d in accordance with the loaded OS (Operating System) program. Similarly, the CPU 10a writes various data stored in the data area 10fb of the auxiliary storage device 10f to the data area 10db of the RAM 10d. The addresses in the RAM 10d where the programs and data are written are then stored in the register 10ac of the CPU 10a.The control unit 10aa of the CPU 10a sequentially reads out these addresses stored in the register 10ac, reads out programs and data from the areas on the RAM 10d indicated by the read addresses, causes the calculation unit 10ab to sequentially execute the calculations indicated by the programs, and stores the calculation results in the register 10ac. With this configuration, the functional configuration of each device is realized.

[0076] The program describing this processing can be recorded on a computer-readable recording medium. Examples of computer-readable recording media are non-transitory recording media. Examples of such recording media include magnetic recording devices, optical disks, magneto-optical recording media, and semiconductor memories.

[0077] The program may be distributed by, for example, selling, transferring, lending, etc. portable recording media such as DVDs and CD-ROMs on which the program is recorded. Furthermore, the program may be stored in a storage device of a server computer, and then transferred from the server computer to other computers via a network, thereby distributing the program.

[0078] A computer that executes such a program may first temporarily store the program recorded on a portable recording medium or transferred from a server computer in its own storage device. Then, when executing a process, the computer reads the program stored on its own recording medium and executes the process in accordance with the read program. Alternatively, the computer may read the program directly from a portable recording medium and execute the process in accordance with the program. Furthermore, the computer may execute the process in accordance with the program each time a program is transferred from a server computer to the computer. Alternatively, the server computer may not transfer the program to the computer, but may instead execute the process through a so-called ASP (Application Service Provider) service, which realizes the processing function by issuing an execution instruction and obtaining the results. Furthermore, the server computer may execute the process at the terminal using a so-called SaaS (Software as a Service) service, which allows users to use part of a server computer along with the program. In this embodiment, the program includes information used for processing by an electronic computer that is equivalent to a program (such as data that is not a direct instruction to a computer but has properties that dictate computer processing).

[0079] Furthermore, in this embodiment, the device is configured by executing a predetermined program on a computer, but at least a part of the processing contents may be realized by hardware.

[0080] [Other Modifications] The present invention is not limited to the above-described embodiment. For example, the contribution of the dataset DS may be calculated or may be obtained by searching a table that associates an index representing the quantity or quality of the dataset DS with the contribution (step S141). Similarly, for example, the data provision fee of the dataset DS may be calculated or may be obtained by searching a table that associates the contribution of the dataset DS with the data provision fee (step S142). Similarly, for example, the usage fee of the CR unit 144 (computational resources) may be calculated or may be obtained by searching a table that associates the performance, type, usage time, learning time, etc. of the CR unit 144 with the usage fee of the CR unit 144 (computational resources) (step S143). Similarly, for example, the remuneration for providing the dataset DS may be calculated or may be obtained by searching a table that associates the data provision fee of the dataset DS with the usage fee of the CR unit 144 and the remuneration for providing the dataset DS (step S144). Furthermore, the various processes described above may not only be executed in chronological order as described, but may also be executed in parallel or individually depending on the processing capacity of the device executing the processes or as necessary. Needless to say, other modifications are possible within the scope of the present invention.

[0081] An outline of an embodiment is described below. [Supplementary Note 1] A federated learning management device having processing circuitry configured to: obtain a contribution of a dataset for federated learning machine learning performed using the dataset, store the contribution of the dataset in memory, and determine a reward amount for the dataset based on at least the contribution of the dataset. [Supplementary Note 2] The federated learning management device of Supplementary Note 1, wherein the processing circuitry is configured to: perform processing for executing the machine learning using a first computational resource; and determine the reward amount based on a usage fee for the first computational resource. [Supplementary Note 3] The federated learning management device of Supplementary Note 1, wherein the processing circuitry is configured to: perform processing for executing the machine learning using a first computational resource when it is determined that a second computational resource is not suitable for the machine learning; and determine the reward amount based on a usage fee for the first computational resource. [Supplementary Note 4] The federated learning management device of Supplementary Note 1, wherein the processing circuit is configured to: perform processing to execute the machine learning using a second computational resource; and, when it is determined that the second computational resource is no longer suitable for the machine learning, switch from the second computational resource to a first computational resource, perform processing to execute the machine learning using the first computational resource, and further determine the remuneration amount based on a usage fee for the first computational resource. [Supplementary Note 5] The federated learning management device of Supplementary Note 4, wherein the processing circuit is further configured to determine the remuneration amount based on the time for which the machine learning is executed using the second computational resource and the time for which the machine learning is executed using the first computational resource. [Supplementary Note 6] The federated learning management device of Supplementary Note 3, wherein the processing circuit is configured to: measure a second partial processing time, which is the processing time required to execute processing including a part of the machine learning using the second computational resource; and determine whether the second computational resource is suitable for the machine learning based on the second partial processing time.[Supplementary Note 7] The federated learning management device of Supplementary Note 4, wherein the processing circuit is configured to measure a second partial processing time, which is the processing time required to execute a process including a part of the machine learning using the second computational resource, and determine whether the second computational resource is suitable for the machine learning based on the second partial processing time. [Supplementary Note 8] A federated learning management method, wherein the processing circuit obtains a contribution of a dataset for machine learning in federated learning performed using the dataset and stores the contribution in memory, and determines a reward amount for the dataset based on at least the contribution of the dataset. [Supplementary Note 9] A non-transitory recording medium storing a program for causing a computer to function as the federated learning management device of Supplementary Note 1.

[0082] DESCRIPTION OF SYMBOLS 11 Federated learning management device 12 User device 13 Data providing device 14 Computational resource providing device 112 CR (computational resource) determination unit 112a Measurement unit 112b Determination unit 112c Test CR unit 114 Learning control unit 114a Measurement unit 114b Evaluation unit 114c CR management unit 115 Settlement unit 115a Contribution level acquisition unit 115b Data provision fee calculation unit 115c Usage fee calculation unit 115d Remuneration amount determination unit

Claims

1. A federated learning management device having a contribution acquisition unit that acquires the contribution of a dataset for machine learning federated learning performed using the dataset, and a reward amount determination unit that determines a reward amount for the dataset based at least on the contribution of the dataset.

2. A federated learning management device according to claim 1, comprising a learning control unit that performs processing to execute the machine learning using a first computational resource, and the reward amount determination unit further determines the reward amount based on the usage fee for the first computational resource.

3. A federated learning management device according to claim 1, comprising a learning control unit that performs processing to execute the machine learning using the first computational resource when it is determined that the second computational resource is not suitable for the machine learning, and the remuneration amount determination unit further determines the remuneration amount based on the usage fee for the first computational resource.

4. A federated learning management device according to claim 1, further comprising a learning control unit that performs processing to execute the machine learning using a second computational resource, and when it is determined that the second computational resource is no longer suitable for the machine learning, switches the second computational resource to a first computational resource and executes the machine learning using the first computational resource, and the remuneration amount determination unit further determines the remuneration amount based on the usage fee for the first computational resource.

5. A federated learning management device according to claim 4, wherein the remuneration amount determination unit further determines the remuneration amount based on the time for which the machine learning was performed using the second computational resource and the time for which the machine learning was performed using the first computational resource.

6. A federated learning management device according to any one of claims 3 to 5, wherein the learning control unit includes: a second measurement unit that measures a second partial processing time, which is the processing time required to execute a process including a part of the machine learning using the second computational resource; and an evaluation unit that determines whether the second computational resource is suitable for the machine learning based on the second partial processing time.

7. A federated learning management method using a federated learning management device, comprising: a contribution acquisition step in which a contribution acquisition unit acquires the contribution of a dataset for machine learning of federated learning performed using the dataset; and a reward amount determination step in which a reward amount determination unit determines a reward amount for the dataset based at least on the contribution of the dataset.

8. A program for causing a computer to function as the associative learning management device of claim 1.

Images