JP6869817B2 - Management equipment, management methods, and programs - Google Patents

Description

The present invention relates to a management device, a management method, and a program for distributing the consideration paid according to the use of the learning model to the contributors.

Conventionally, a learning model for predicting output data with respect to input data has been created based on machine learning (see, for example, Patent Document 1).

Japanese Unexamined Patent Publication No. 2015-060237

The learning model has a plurality of layers created by learning using a neural network. In the creation of a learning model, a learning model may be created by adding a new layer to a learning model created by another person. In this case, the learning model is created by contributors as multiple authors.

By the way, it is required to appropriately pay the contributor for the use of the learning model. However, traditionally, it has not been considered to properly reward contributors for the use of learning models.

Therefore, the present invention has been made in view of these points, and provides a management device, a management method, and a program capable of appropriately paying a reward to a contributor according to the use of a learning model. With the goal.

The management device according to the first aspect of the present invention has a storage unit that holds a learning model created by one or more contributors in association with a contribution rate of each contributor regarding the creation of the learning model, via a network. A reception unit that receives a request for using the learning model from another device, and a case where the consideration paid in return for using the learning model by the other device is distributed to each contributor according to the contribution rate. It is equipped with a reward calculation unit for calculating rewards.

The learning model is the weight of each of a plurality of layers generated by learning using a neural network, and the management device calculates the contribution ratio based on the ratio of the number of layers generated by each contributor to the total number of layers. A contribution rate management unit may be further provided.

The contribution rate management unit may calculate the contribution rate based on the amount of data used by the contributor to train the neural network to generate the learning model.
The contribution rate management unit may update the contribution rate when the learning model is updated.

The management device further includes a hash generation unit that generates a data string having a hash value of the weight of each layer of the learning model as an element, and the contribution rate management unit further contributes when the data string is changed. You may update the rate.
The management device may further include a quantization unit that quantizes the weights of the respective layers, and the hash generation unit may calculate a hash value of the weights quantized by the quantization unit.

In the management method according to the second aspect of the present invention, the processor holds the contribution rate data in which the learning model created by one or more contributors and the contribution rate of each contributor regarding the creation of the learning model are associated with each other. The step of reading the contribution rate data from the storage unit, the step of accepting the usage request of the learning model from another device via the network, and the consideration paid in return for the use of the learning model by the other device. , And the step of calculating each reward when distributing to each contributor according to the contribution rate.

The program according to the third aspect of the present invention stores a computer in which a learning model created by one or more contributors and a contribution rate data in which the contribution rate of each contributor with respect to the creation of the learning model is associated with each other. A function of reading the contribution rate data from the unit, a function of accepting a request for using the learning model from another device via a network, and a consideration paid in return for using the learning model by the other device. The function of calculating each reward when distributing to each contributor according to the contribution rate is realized.

According to the present invention, there is an effect that the contributor can be appropriately paid according to the use of the learning model.

It is a figure which shows the outline of the management apparatus which concerns on this embodiment. It is a figure which shows the structure of the management apparatus which concerns on this embodiment. It is a figure which shows an example of the learning model information which concerns on this embodiment. It is a figure which shows an example of the license information which concerns on this embodiment. It is a figure which shows the structure of the model registration part which concerns on this embodiment. It is a flowchart which shows the flow of the process which concerns on registration or update of the learning model which concerns on this Embodiment. It is a flowchart which shows the flow of the process which concerns on use of the learning model which concerns on this embodiment, and the calculation of a reward.

[Overview of management device 1]
FIG. 1 is a diagram showing an outline of the management device 1 according to the present embodiment. The management device 1 is a device that accepts the registration of the learning model of the neural network and accepts the use of the registered learning model for a fee. The management device 1 is communicably connected to the contributor terminal 2 and the user terminal 3 via a communication network such as a LAN (Local Area Network) or the Internet.

In explaining this embodiment, first, a neural network will be briefly described. The basic configuration of a neural network is represented by a superposition (or graph structure) of a plurality of types of layers.

FIG. 1 shows a feedforward neural network. A feedforward neural network is composed of three layers, an input layer, a hidden layer, and an output layer, and is unidirectional from the input layer to the output layer. Propagate. The hidden layer can be composed of a plurality of layers in a graph shape. Each layer has multiple units (neurons). In each layer, the parameter of the function that connects the unit in the front layer to the unit in the rear layer is called "weight". Learning in this embodiment is to calculate an appropriate "weight" as a parameter of this function. Further, the learning model in the present embodiment is the weight of each of the plurality of layers and is used to realize the task.

Learning in a neural network means that the weight of each layer is optimally updated by using the error between the output value from the output layer and the label of the training data with respect to the training data. A "loss function" is defined to calculate the error. The error is propagated one after another from the output layer side to the input layer side by the "error backpropagation method", and the weight of each layer is updated little by little. A learning model is created by performing a convergence calculation that adjusts the weights of each layer to the appropriate values so that the error is finally small.

By the way, a user who creates a learning model may create a learning model by updating an existing weight by re-learning a learning model created in advance or by adding a new layer. In this case, a plurality of users are contributing to the creation of the learning model. Hereinafter, the user who contributed to the creation of the learning model is referred to as a "contributor".

In the example shown in FIG. 1, the learning model M is generated. The learning model M includes weights W1 to Wn. Here, as shown in FIG. 1, the contributor A creates the input layer and the weights W1 to Wi corresponding to a part of the hidden layer, and the contributor B creates the weights Wi + 1 to Wn corresponding to the other part of the hidden layer. Created by. In this case, the contributor A and the contributor B contribute to the creation of the learning model M.

The management device 1 according to the present embodiment associates the learning model with the contribution rate of each contributor to the creation of the learning model in order to appropriately distribute the consideration paid in return for using the learning model to each contributor. And memorize it ((1) in FIG. 1).

Further, the management device 1 accepts the license registration of the learning model from the user terminal 3 used by the user who wants to use the learning model ((2) in FIG. 1). Here, the license is managed by associating a model ID as model identification information for identifying the learning model with a user ID for identifying the user of the user who uses the learning model. The model identification information is generated based on the hash values of one or more subsets obtained by partitioning the training model. By doing so, since the learning model created by using the other learning model includes the model ID corresponding to the other learning model, the management device 1 is one or a plurality of contributors who created the learning model. Can be identified.

After that, the management device 1 receives a request for using the learning model ((3) in FIG. 1). When the management device 1 receives the usage request, the management device 1 authenticates whether or not the user can use the learning model. When the authentication is successful, the management device 1 executes a task using the learning model and outputs the execution result to the user terminal 3 ((4) in FIG. 1). The management device 1 calculates the consideration for using the learning model, and charges the user of the user terminal 3 for payment of the consideration. After that, the management device 1 calculates the reward to be paid to each contributor based on the consideration paid by the user and the contribution rate associated with the learning model ((5) in FIG. 1). By doing so, the management device 1 can appropriately pay the contributor a reward according to the use of the learning model.
Hereinafter, the configuration of the management device 1 will be described.

[Configuration example of management device 1]
FIG. 2 is a diagram showing a configuration of the management device 1 according to the present embodiment.
The management device 1 includes a storage unit 11 and a control unit 12.

The storage unit 11 is, for example, a ROM (Read Only Memory), a RAM (Random Access Memory), or the like. The storage unit 11 stores various programs for operating the management device 1. For example, the storage unit 11 uses the control unit 12 of the management device 1 as a model registration unit 13, a license registration unit 14, a use reception unit 15, an authentication unit 16, a learning execution unit 17, a consideration calculation unit 18, and a reward calculation unit, which will be described later. Store the management program that functions as unit 19. Further, the storage unit 11 stores the learning model information including the learning model and the license information of the learning model.

FIG. 3 is a diagram showing an example of learning model information according to the present embodiment. As shown in FIG. 3, the learning model information includes a registrant, a model name, a description, a learning model, a model ID, a post-quantized data string, a contributor, and a contribution rate.

The registrant indicates the contributor who registered the learning model, and the contributor ID that identifies the contributor is stored. The model name is information for identifying the model. In this embodiment, the model name is different for each of all learning models.

The descriptive text is a character string for explaining the contents of the model. The model ID is information that identifies the learning model. The post-quantization data sequence is a hash value after the model ID is quantized. The contributor indicates the contributor who contributed to the creation of the learning model, and the contributor ID is stored. Here, the contributor and the contribution rate are associated with one or more for one learning model.

FIG. 4 is a diagram showing an example of license information according to the present embodiment. As shown in FIG. 4, the license information includes a user ID of a user who uses the learning model, a model name, and a model ID.

Returning to FIG. 2, the control unit 12 is, for example, a CPU (Central Processing Unit). The control unit 12 controls the functions related to the management device 1 by executing various programs stored in the storage unit 11. By executing the management program, the control unit 12 functions as a model registration unit 13, a license registration unit 14, a usage reception unit 15, an authentication unit 16, a learning execution unit 17, a consideration calculation unit 18, and a reward calculation unit 19. ..

[Registration and update of learning model]
The model registration unit 13 accepts new registration or update of the learning model from the contributor terminal 2 used by the contributor. FIG. 5 is a diagram showing a configuration of a model registration unit 13 according to the present embodiment. As shown in FIG. 5, the model registration unit 13 includes an acquisition unit 131, a division unit 132, a first hash generation unit 133, a quantization unit 134, a second hash generation unit 135, and a contribution rate management unit 136. And a memory control unit 137.

The acquisition unit 131 acquires the learning model from the contributor terminal 2. Specifically, when accepting a new registration of a learning model, the acquisition unit 131 receives a learning model, a learning model name, a description of the learning model, a label of the learning model, and a learning model from the contributor terminal 2. Acquires the contributor ID of the contributor who created. Here, the learning model includes an index assigned to each of the plurality of layers and a plurality of weight values indicating the weights of the plurality of layers.

Further, when accepting the update of the learning model, the acquisition unit 131 first acquires the contributor ID from the contributor terminal 2. The acquisition unit 131 refers to the learning model information stored in the storage unit 11, identifies the model name associated with the contributor ID, and accepts the selection of any one of the model names. Then, the acquisition unit 131 acquires the updated learning model from the contributor terminal 2. The acquisition unit 131 may acquire the label of the learning model even when the learning model is updated.

The division unit 132 divides the learning model into one or a plurality of subsets. The dividing unit 132 divides a plurality of layers and sets the weights corresponding to each layer as a subset. Here, the division unit 132 generates a subset including the label of the learning model acquired by the acquisition unit 131 together with the learning model.

The first hash generation unit 133 generates a hash value for each of the one or a plurality of subsets generated by the division unit 132 to generate a model ID for specifying the learning model. Specifically, first, the first hash generation unit 133 generates a hash value of each of the plurality of weight values by applying a hash function to each of the plurality of weight values indicating one or a plurality of subsets. To do. Then, the first hash generation unit 133 concatenates the hash values generated for each of the plurality of weight values in the order of the indexes assigned to each layer, so that the first hash generation unit 133 is a data string having these hash values as elements. Is generated as a model ID.

The quantization unit 134 quantizes each of the subsets divided by the division unit 132, which is the weight of each layer of the learning model. For example, the quantization unit 134 divides a range in which a weight value can be taken into a plurality of ranges, and sets a representative value in each of the plurality of ranges. The quantization unit 134 quantizes a subset by converting a plurality of weight values into representative values corresponding to the range to which the weight values belong. By doing so, the management device 1 can reduce the processing load without changing the contribution rate when the learning model is finely updated.

The second hash generation unit 135 generates a post-quantization data string having the hash value of the weight of each layer of the learning model as an element. Specifically, first, the second hash generation unit 135 applies a hash function to each of the plurality of weight values indicating the subset indicated by the quantized unit 134, so that each of the plurality of weight values can be obtained. Calculate the hash value. Then, the second hash generation unit 135 concatenates the hash values generated for each of the plurality of weight values in the order of the indexes assigned to each layer, so that the post-quantization data having these hash values as elements is used. Generate a column.

The contribution rate management unit 136 calculates the contribution rate of each contributor regarding the creation of the learning model acquired by the acquisition unit 131. When newly registering a learning model, the contribution rate management unit 136 refers to the learning model information stored in the storage unit 11 and uses the model ID generated by the first hash generation unit 133 as the learning model information. It is determined whether or not the model ID of another stored learning model is included.

When the model ID of another learning model is not included, the contribution rate management unit 136 determines that the newly registered learning model is originally created by the contributor who registers, and determines the contribution rate of the contributor. Calculate as 1 (100%).

When the model ID of another learning model is included, the contribution rate management unit 136 refers to the other learning model included in the learning model information stored in the storage unit 11 and refers to the other learning model. Identify the number of layers in the learning model. Then, the contribution rate management unit 136 subtracts the number of layers of the other learning model from the total number of layers of the newly registered learning model, and specifies the number of layers generated by the contributor who performs the registration.

Further, the contribution rate management unit 136 refers to the learning model information stored in the storage unit 11, and based on the contributor and contribution rate of the other learning model and the number of layers of the other learning model, the said Identify the number of layers generated by each contributor who generated other learning models. Then, the contribution rate management unit 136 calculates the contribution rate based on the ratio of the number of layers generated by each contributor to the total number of layers of the newly registered learning model.

Further, the contribution rate management unit 136 updates the contribution rate when the learning model is updated. Specifically, the contribution rate management unit 136 refers to the learning model information stored in the storage unit 11 and specifies the post-quantized data string of the learning model to be updated. Then, the contribution rate management unit 136 is in the case where the specified post-quantization data string and the post-quantization data string generated by the second hash generation unit 135 are different, that is, when the post-quantization data string is changed. In addition, it is determined that the learning model has been updated, and the contribution rate is calculated. Since the method of calculating the contribution rate when the learning model is updated is the same as the method of calculating the contribution rate when the learning model is newly registered, the description thereof will be omitted.

The contribution rate management unit 136 may calculate the contribution rate based on the amount of data used for learning the neural network in order for each contributor to generate a learning model. In this case, when the learning model is registered, the acquisition unit 131 acquires information indicating the amount of data used for learning the neural network in order to generate the learning model, and the learning model information stored in the storage unit 11 Stores information indicating the amount of data.

Further, the contribution rate management unit 136 may calculate the contribution rate based on the total cost of the computational resources used for learning. In this case, when the learning model is registered, the acquisition unit 131 acquires computational resource information indicating the amount of computational resources consumed for learning the neural network in order to generate the learning model, and is stored in the storage unit 11. The computational resource information is stored in the learning model information. Here, the resource consumption information includes information indicating the CPU usage time and information indicating the usage area (memory usage amount) of the storage unit 11.

Further, in order to calculate the contribution rate based on the computational resources, the management device 1 may include a model generation unit that generates a learning model by learning the neural network. Then, the model generation unit receives the data used for learning the neural network from the contributor terminal 2 and learns the neural network. Further, the model generation unit stores the contributor ID, the model ID of the generated learning model, and the consumption resource information indicating the computational resources consumed in the learning in association with each other in the storage unit 11. The contribution rate management unit 136 identifies the ratio of the computational resources used by each contributor to generate the learning model based on the computational resources indicated by the consumption resource information stored in the storage unit 11, and contributes based on the ratio. Calculate the rate.

The contribution rate management unit 136 refers to the learning model information and identifies the information indicating the amount of data corresponding to the learning model and the information indicating the amount of data corresponding to other learning models included in the learning model. .. Then, the contribution rate management unit 136 calculates the contribution rate based on the ratio of the amount of data corresponding to the learning model and the amount of data corresponding to other learning models. In the generation of the learning model, the larger the amount of data used for learning, the higher the learning effect and the more useful the learning model. Therefore, the management device 1 can set a high contribution rate of the contributor who has contributed to the improvement of the usefulness of the learning model by calculating the contribution rate based on the amount of data used for learning.

The memory control unit 137 stores the learning model in the storage unit 11. The memory control unit 137 identifies the contributor ID acquired by the acquisition unit 131 as a registrant. When the memory control unit 137 accepts the new registration of the learning model, the specified registrant, the learning model acquired by the acquisition unit 131, the learning model name, the description and the label, and the first hash generation unit 133 The generated model ID, the post-quantization data string generated by the second hash generation unit 135, and the contributor and contribution rate specified by the contribution rate management unit 136 are associated and stored in the storage unit 11 as learning model information. ..

Further, when the storage control unit 137 accepts the update of the learning model, when the contribution rate management unit 136 determines that the learning model has been updated based on the data string after quantization, the learning stored in the learning model information is stored. Update the model, model ID, post-quantized data sequence, contributors and contribution rates. The storage control unit 137 includes a learning model acquired by the acquisition unit 131, a model ID generated by the first hash generation unit 133, a quantized data string generated by the second hash generation unit 135, and a contribution rate management unit 136. Update the learning model information based on the contributors and contribution rates identified by.

Further, when the storage control unit 137 receives the update of the learning model, if the contribution rate management unit 136 determines that the learning model has not been updated based on the data string after quantization, it is stored in the learning model information. The learning model and the model ID are updated to the learning model acquired by the acquisition unit 131 and the model ID generated by the first hash generation unit 133.

[Registration of license information]
The explanation is returned to FIG. The license registration unit 14 accepts license registration from the user of the user terminal 3. For example, the license registration unit 14 causes the user terminal 3 to display the model name, the model ID, and the description, and accepts the selection of the learning model to be used by the user. Here, the license registration unit 14 may calculate the license fee of the learning model based on, for example, the total number of layers of the learning model and the label, and display the license fee on the user terminal 3.

When the learning model is selected, the user terminal 3 transmits a license registration request including the model name, the model ID, and the user ID that identifies the user to the management device 1. When the license registration unit 14 receives the license registration request, the license registration unit 14 stores the user ID, the model name, and the model ID included in the license registration request in the storage unit 11 as license information in association with each other.

[Use of learning model]
The use reception unit 15 receives a request for use of the learning model from the user terminal 3 as another device via the communication network. Specifically, the usage reception unit 15 receives the usage request of the learning model from the user terminal 3 by receiving the usage request of the learning model including the model name of the learning model used by the user and the user ID. .. In the present embodiment, the usage reception unit 15 receives the usage request of the learning model from the user terminal 3, but the present invention is not limited to this. The use reception unit 15 may receive a use request from the contributor terminal 2.

When the use receiving unit 15 receives the learning model usage request from the user terminal 3, the authentication unit 16 refers to the license information stored in the storage unit 11, and the user of the user terminal 3 uses the learning model. Determine if it is possible.

Specifically, the authentication unit 16 allows the user of the user terminal 3 to store the model name of the learning model included in the usage request of the learning model and the user ID in the license information in association with each other. , Judge that the learning model can be used.

Further, when the model name of the learning model included in the usage request of the learning model and the user ID are associated with each other and stored in the license information, the authentication unit 16 allows the user of the user terminal 3 to perform the learning. Determine that the model is unusable. When the authentication unit 16 determines that the learning model cannot be used, the authentication unit 16 notifies the user terminal 3 of a message indicating that the learning model cannot be used.

When the user can use the learning model, the learning execution unit 17 outputs the execution result of the learning model regarding the data for executing the task acquired from the user terminal 3. Specifically, first, when the authentication unit 16 determines that the user of the user terminal 3 can use the learning model, the learning execution unit 17 acquires one or more data for executing the task from the user terminal 3. .. The learning execution unit 17 uses the data acquired from the user terminal 3 as input data, and acquires the execution result by executing the task using the learning model corresponding to the usage request. The learning execution unit 17 outputs the acquired execution result to the user terminal 3.

[Calculation of compensation for using the learning model and reward for each contributor]
The consideration calculation unit 18 calculates the amount to be charged to the user based on the usage mode of the learning execution unit 17 for executing the task. For example, the consideration calculation unit 18 calculates the billing amount to be charged to the user based on the usage load of the learning execution unit 17 for executing the task. Here, the usage load of the learning execution unit 17 is the time spent by the learning execution unit 17, that is, the CPU usage time when the task based on the learning model used by the user is executed. By doing so, the management device 1 can fairly calculate the billing amount from the viewpoint of the usage load.

The consideration calculation unit 18 may calculate the billing amount to be charged to the user based on the occupancy rate of the learning execution unit 17 for executing the task. Here, the occupancy rate of the learning execution unit 17 is the CPU utilization rate of the learning execution unit 17 when the task based on the learning model used by the user is executed. By doing so, the management device 1 can fairly calculate the billing amount based on the CPU utilization rate when used by a plurality of users.

The consideration calculation unit 18 notifies the user terminal 3 of the calculated billing amount. The user of the user terminal 3 pays the notified billing amount to the administrator of the management device 1 by a predetermined method.

The reward calculation unit 19 calculates each reward when the consideration paid in return for the use of the learning model by the user terminal 3 is distributed to each contributor according to the contribution rate related to the creation of the learning model.

Specifically, first, the reward calculation unit 19 refers to the learning model information stored in the storage unit 11 and specifies the contribution rate of each contributor of the learning model for which the consideration has been calculated. Then, the reward calculation unit 19 calculates the reward of each contributor by multiplying the consideration calculated by the consideration calculation unit 18 with the contribution rate of each specified contributor.

Here, the remuneration calculation unit 19 may calculate the remuneration of the administrator of the management device 1 and subtract the remuneration of the administrator from the remuneration calculated by the consideration calculation unit 18. Then, the reward calculation unit 19 may calculate the reward of each contributor by multiplying the consideration after subtraction by the contribution rate of each specified contributor. The remuneration calculation unit 19 pays the calculated billing amount to each contributor by a predetermined method.

[Process flow related to learning model registration]
Subsequently, the flow of processing in the management device 1 will be described. First, the flow of processing related to registration or update of the learning model will be described. FIG. 6 is a flowchart showing a flow of processing related to registration or update of the learning model according to the present embodiment.

First, the acquisition unit 131 acquires the learning model (S10).
Subsequently, the division unit 132 divides the learning model to generate a subset (S20).

Subsequently, the first hash generation unit 133 generates a hash value of the subset generated in S20, and generates a model ID of the learning model based on the hash value (S30).
Subsequently, the quantization unit 134 quantizes the subset (S40). Then, the second hash generation unit 135 generates a hash value of the quantized subset, and generates a quantized data string based on the hash value (S50).

Subsequently, the contribution rate management unit 136 calculates the contribution rate of each contributor regarding the creation of the learning model (S60).
Subsequently, the memory control unit 137 updates the learning model information by registering or updating the learning model (S70).

[Flow of processing related to use of learning model and reward calculation]
Next, the flow of processing related to the use of the learning model and the calculation of the reward will be described. FIG. 7 is a flowchart showing the flow of processing related to the use of the learning model and the calculation of the reward according to the present embodiment.

First, the use reception unit 15 receives a request for use of the learning model (S110). Here, the request for using the learning model includes the model name and the user ID of the learning model.
Subsequently, the authentication unit 16 authenticates whether or not the learning model for which the usage request has been accepted can be used (S120).

Subsequently, the learning execution unit 17 acquires input data for executing a task using the learning model according to the success of the authentication in the authentication unit 16 (S130).
Subsequently, the learning execution unit 17 executes a task using the acquired input data and the learning model corresponding to the usage request (S140). As a result, the learning execution unit 17 acquires the execution result of the task.

Subsequently, the consideration calculation unit 18 calculates the billing amount related to the use of the learning model based on the usage mode of the learning model (S150).
Subsequently, the reward calculation unit 19 corresponds to the learning model used and calculates the reward for each contributor based on the contribution rate stored in the learning model information (S160).

[Effect in this embodiment]
As described above, the management device 1 according to the present embodiment stores the learning model created by one or more contributors in association with the contribution rate related to the creation of the learning model of each contributor, and stores the user terminal 3 via the network. The request for use of the learning model is received from, and the compensation paid in return for the use of the learning model by the user terminal 3 is calculated for each contributor when the consideration is distributed to each contributor according to the contribution rate corresponding to the learning model.

By doing so, the management device 1 can appropriately pay a reward to the contributor who contributed to the creation of the learning model. In particular, when the learning model is created by a plurality of contributors, the reward of each contributor can be fairly calculated based on the contribution rate of each contributor regarding the creation of the learning model.

Further, the management device 1 according to the present embodiment divides the learning model into one or a plurality of subsets, generates hash values for each subset, generates a model ID for specifying the learning model, and generates a model ID. And the license information associated with the user ID that can use the learning model is stored. By doing so, the learning model created by using at least a part of the other learning model includes the model ID corresponding to the other learning model. Therefore, the management device 1 can manage whether the learning model is created by using another learning model, and can identify one or a plurality of contributors who created the learning model.

Although the present invention has been described above using the embodiments, the technical scope of the present invention is not limited to the scope described in the above embodiments. It will be apparent to those skilled in the art that various changes or improvements can be made to the above embodiments. Further, in particular, the specific embodiments of the distribution / integration of the devices are not limited to those shown above, and all or a part thereof may be arbitrarily added according to various additions or functional loads. It can be functionally or physically distributed / integrated in units.

For example, in the above-described embodiment, the management device 1 accesses each of the plurality of communication devices 4, and acquires the logs generated by each of the plurality of communication devices 4, but the present invention is not limited to this. For example, a log acquisition device for acquiring logs generated by each of the plurality of communication devices 4 and a management device 1 are provided, and the management device 1 is generated by each of the plurality of communication devices 4 from the log acquisition device. You may get the log.

For example, in the above-described embodiment, the case where the contribution rate management unit 136 calculates the contribution rate based on the ratio of the number of layers generated by each contributor to the total number of layers of the newly registered learning model has been described. Instead, the contribution rate management unit 136 may calculate the contribution rate based on the ratio of the weights included in the layers generated by each contributor to the total amount of data of the weights provided in each layer of the newly registered learning model. Good. Since the amount of weight data provided by each layer may differ depending on the layer, the amount of weight data actually generated by each contributor can be reflected in the contribution rate.

1 ... management device, 11 ... storage unit, 12 ... control unit, 13 ... model registration unit, 131 ... acquisition unit, 132 ... division unit, 133 ... first hash Generation unit, 134 ... Quantization unit, 135 ... Second hash generation unit, 136 ... Contribution rate management unit, 137 ... Memory control unit, 14 ... License registration unit, 15 ... Usage reception department, 16 ... authentication department, 17 ... learning execution department, 18 ... consideration calculation department, 19 ... reward calculation department, 2 ... contributor terminal, 3 ... user terminal

Claims (9)

For a learning model of a neural network having a plurality of layers created by one or a plurality of contributors, the contribution ratio of each contributor to the creation of the learning model is calculated based on the ratio of the number of layers generated by each contributor to the total number of layers. Contribution rate management department to calculate and
A storage unit that associates and holds the learning model with the contribution rate of each contributor calculated by the contribution rate management unit.
A reception unit that accepts requests to use the learning model from other devices via the network,
A reward calculation unit that calculates each reward when the consideration paid in return for the use of the learning model by the other device is distributed to each contributor according to the contribution rate stored in the storage unit.
A management device equipped with. For a learning model of a neural network having a plurality of layers created by one or a plurality of contributors, the said of each contributor is based on the amount of data used by each contributor to train the neural network to generate the learning model. Contribution rate management department that calculates the contribution rate for creating a learning model,
A storage unit that associates and holds the learning model with the contribution rate of each contributor calculated by the contribution rate management unit.
A reception unit that accepts requests to use the learning model from other devices via the network,
A reward calculation unit that calculates each reward when the consideration paid in return for the use of the learning model by the other device is distributed to each contributor according to the contribution rate stored in the storage unit.
A management device equipped with. It also has an update reception unit that accepts updates of the learning model.
The contribution rate management unit updates the contribution rate when the update reception unit receives the update of the learning model.
The management device according to claim 1 or 2. Further, a hash generation unit for generating a data string having the hash value of the weight of each layer of the learning model as an element is provided.
The contribution rate management unit updates the contribution rate when the data string is changed.
The management device according to any one of claims 1 to 3. The possible range of each weight of each layer is divided into a plurality of ranges, and each weight of each layer is quantized by converting it into a representative value set for the range to which each weight of each layer belongs. With a further quantization unit
The hash generation unit calculates the hash value of the weight by calculating the hash value of the representative value converted from the weight by the quantization unit.
The management device according to claim 4. The processor,
For a learning model of a neural network having a plurality of layers created by one or a plurality of contributors, the contribution ratio of each contributor to the creation of the learning model is calculated based on the ratio of the number of layers generated by each contributor to the total number of layers. Steps to calculate and
A step of storing the contribution rate data in which the learning model is associated with the contribution rate of each contributor in the storage unit, and
A step of reading the contribution rate data from the storage unit,
A step of accepting a request for using the learning model from another device via a network,
A step of calculating each reward when the consideration paid in return for the use of the learning model by the other device is distributed to each contributor according to the contribution rate.
How to manage to run. The processor,
For a learning model of a neural network having a plurality of layers created by one or a plurality of contributors, the said of each contributor based on the amount of data used by each contributor to train the neural network to generate the learning model. Steps to calculate the contribution rate for creating a learning model,
A step of storing the contribution rate data in which the learning model is associated with the contribution rate of each contributor in the storage unit, and
A step of reading the contribution rate data from the storage unit,
A step of accepting a request for using the learning model from another device via a network,
A step of calculating each reward when the consideration paid in return for the use of the learning model by the other device is distributed to each contributor according to the contribution rate stored in the storage unit.
How to manage to run. On the computer
For a learning model of a neural network having a plurality of layers created by one or a plurality of contributors, the contribution ratio of each contributor to the creation of the learning model is calculated based on the ratio of the number of layers generated by each contributor to the total number of layers. Function to calculate and
A function of storing the contribution rate data in which the learning model is associated with the contribution rate of each contributor in the storage unit, and
A function of reading the contribution rate data from the storage unit and
A function that accepts usage requests for the learning model from other devices via the network,
A function of calculating each reward when the consideration paid in return for the use of the learning model by the other device is distributed to each contributor according to the contribution rate stored in the storage unit.
A program that realizes. On the computer
For a learning model of a neural network having a plurality of layers created by one or a plurality of contributors, the said of each contributor is based on the amount of data used by each contributor to train the neural network to generate the learning model. A function to calculate the contribution rate for creating a learning model, and
A function of storing the contribution rate data in which the learning model is associated with the contribution rate of each contributor in the storage unit, and
A function of reading the contribution rate data from the storage unit and
A function that accepts usage requests for the learning model from other devices via the network,
A function of calculating each reward when the consideration paid in return for the use of the learning model by the other device is distributed to each contributor according to the contribution rate stored in the storage unit.
A program that realizes.

Images