Management device, management method, and management program

Abstract

To provide an appropriate learning model.SOLUTION: A management device includes: a storage unit configured to store a plurality of first learning models learned for each group ID for specifying an attribute group that is a combination of a plurality of attributes; a reception unit configured to receive group information including the group ID or the attribute group; a specification unit configured to specify a first learning parameter of the first learning model corresponding to the group ID based on the group information received by the reception unit; and a transmission unit configured to transmit the first learning parameter specified by the specification unit to a transmission source of the group information.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to a management device, a management method, and a management program for managing data. [Background technology]

[0002] Federated learning (for example, see Patent Document 1 below) deploys a learning model, which is a trained AI (Artificial Intelligence), on-site and provides it to users. Federated learning also includes a technology in which multiple learning models are sent to the site from a server in advance, and the data is clustered on a computer at the site to train the learning model (for example, see Patent Document 2 below).

[0003] Patent document 1 discloses a method including the steps of acquiring resource information of the multiple terminal devices for each task in a federated learning system that includes a server and multiple terminal devices and is used to accomplish multiple tasks, determining a target terminal device corresponding to the task using the resource information, and training a global model corresponding to the task by the target terminal device until the global model satisfies preset conditions.

[0004] Patent document 2 discloses an information processing device including a learning unit that clusters hierarchical data based on a plurality of inference models distributed from an information processing server and performs learning using the inference model corresponding to each cluster, and a communication unit that transmits intermediate results generated for each cluster in the learning by the learning unit to the information processing server, wherein the hierarchical data includes information identifying main elements and logs collected or generated in association with the main elements. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] JP 2022-137182 A [Patent Document 2] International Publication No. WO2022 / 239264 Summary of the Invention [Problem to be solved by the invention]

[0006] In Patent Document 1, learning models trained at various sites are collected in an aggregation server such as a cloud to generate a learning model, and the trained learning model is provided to users who want a new learning model. The technology in Patent Document 1 can prevent the leakage of private data, but it is not designed to provide optimal learning models that meet the needs of various sites in various industries.

[0007] In Patent Document 2, a plurality of learning models that take into account differences in data are prepared in an aggregation server such as a cloud. The technology in Patent Document 2 is a system that provides a learning model for a specific industry, and is not designed to provide an appropriate learning model from learning models for various industries.

[0008] The present invention aims to provide an appropriate learning model. [Means for solving the problem]

[0009] A management device according to one aspect of the invention disclosed in the present application is characterized in that it has a memory unit that stores multiple first learning models learned for each group ID that identifies an attribute group, which is a combination of multiple attributes, a receiving unit that receives group information including the group ID or an attribute group, an identification unit that identifies a first learning parameter of the first learning model corresponding to the group ID based on the group information received by the receiving unit, and a transmitting unit that transmits the first learning parameter identified by the identification unit to a sender of the group information. Effect of the Invention

[0010] According to the representative embodiment of the present invention, it is possible to provide an appropriate learning model. Problems, configurations and effects other than those described above will become apparent from the following description of the embodiments. [Brief description of the drawings]

[0011] [Figure 1] FIG. 1 is an explanatory diagram showing an example of the system configuration of a federated learning system. [Diagram 2] FIG. 2 is a block diagram showing an example of the hardware configuration of a computer. [Diagram 3] FIG. 3 is a block diagram illustrating an example of a functional configuration of the server. [Figure 4] FIG. 4 is an explanatory diagram showing an example of the i-th learning model DB set. [Diagram 5] FIG. 5 is an explanatory diagram illustrating an example of the group management table. [Figure 6] FIG. 6 is a block diagram illustrating an example of a functional configuration of the learning unit. [Figure 7] FIG. 7 is a block diagram illustrating an example of a functional configuration of an edge. [Figure 8] FIG. 8 is a sequence diagram showing an example of a first learning sequence in the federated learning system. [Figure 9] FIG. 9 is a sequence diagram showing an example of a second learning sequence in the federated learning system. [Figure 10] FIG. 10 is a sequence diagram showing an example of a third learning sequence in the federated learning system. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] <Associative Learning System> 1 is an explanatory diagram showing an example of the system configuration of a federated learning system. The federated learning system 100 has a server 101 and edges 102-1 to 102-f (f is an integer equal to or greater than 1). When there is no need to distinguish between the edges 102-1 to 102-f, they are referred to as edges 102. The server 101 and the edges 102 are connected to each other so as to be able to communicate with each other via a network 110 such as the Internet, a LAN (Local Area Network), or a WAN (Wide Area Network).

[0013] Edges 102-1 to 102-f are installed in sites ST1 to STf, respectively. When edges 102-1 to 102-f are not distinguished, they are written as edge 102. Edge 102 has one or more learning models. A learning model in edge 102 is referred to as a local learning model. When sites ST1 to STf are not distinguished, they are written as site ST. Sites ST1 to STf are communicatively connected to sensors 131-1 to 131-f and terminals 132-1 to 132-f installed in the work area via LANs 120-1 to 120-f.

[0014] When sensors 131-1 to 131-f are not differentiated from one another, they are referred to as sensors 131. When terminals 132-1 to 132-f are not differentiated from one another, they are referred to as terminals 132. When LANs 120-1 to 120-f are not differentiated from one another, they are referred to as LANs 120.

[0015] The work area is an area in the site ST where a specific work is performed, and various sensors 131 such as a surveillance camera, an infrared sensor, a vibration sensor, and an acceleration sensor are installed according to the specific work. Data detected by these sensors 131 is transmitted to the edge 102 via the LAN 120. Data from the terminal 132 is also transmitted to the edge 102 via the LAN 120.

[0016] For example, the sites ST1 and ST2 are construction sites, and various sensors 131 are installed in a heavy machinery work area and a human work area. The site STf is a manufacturing site, and various sensors 131 are installed in an assembly manufacturing area and a human work area.

[0017] The sites ST1 to STf have various industries, uses, varieties, and regions. The parameters of the federated learning are different depending on the industry, use, variety, and region of the site ST. For example, as described above, the industries of the sites ST1 and ST2 are the construction industry, and the site STf is the assembly manufacturing industry. Therefore, it is necessary to distinguish the local learning model for each industry.

[0018] In addition, even in the construction industry, there are different types of buildings, detached houses, etc. Therefore, it is necessary to distinguish the local learning model according to the type of building.

[0019] An application is a method or purpose of using local learning, and it is diverse, including failure sign detection, work plan formulation, and personal authentication. Therefore, it is necessary to distinguish local learning models according to the application. In addition, since the climate and human behavior differ depending on the region where the site ST is installed, it is also necessary to distinguish local learning models according to the region. In addition, since the production volume, number of workers, and number of machine tools at the site ST differ depending on the scale of the site ST, it is also necessary to distinguish local learning models according to the scale.

[0020] The server 101 is a management device that learns an optimal learning model according to the industry, purpose, product type, region, and scale of such sites ST1 to STf. The industry, purpose, product type, region, and scale are collectively referred to as an attribute group. The site ST has at least one attribute group. For example, if there are learning model groups for multiple uses in the site ST, the site ST belongs to attribute groups equal to the number of types of uses. The industry, purpose, product type, region, and scale that constitute the attribute group are merely examples, and other attributes may be used.

[0021] <Example of computer hardware configuration> FIG. 2 is a block diagram showing an example of a hardware configuration of a computer (server 101, edge 102, terminal 132). The computer 200 has a processor 201, a storage device 202, an input device 203, an output device 204, and a communication interface (communication IF) 205. The processor 201, the storage device 202, the input device 203, the output device 204, and the communication IF 205 are connected by a bus 206. The processor 201 controls the computer 200. The storage device 202 is a working area for the processor 201. The storage device 202 is a non-transient or temporary recording medium that stores various programs and data. Examples of the storage device 202 include a ROM (Read Only Memory), a RAM (Random Access Memory), a HDD (Hard Disk Drive), and a flash memory. The input device 203 inputs data. Examples of the input device 203 include a keyboard, a mouse, a touch panel, a numeric keypad, a scanner, a microphone, and a sensor. The output device 204 outputs data. Examples of the output device 204 include a display, a printer, and a speaker. The communication IF 205 connects to the networks 110 and 120 and transmits and receives data.

[0022] <Server 101> 3 is a block diagram showing an example of a functional configuration of the server 101. The server 101 includes a receiving unit 301, a transmitting unit 302, an identifying unit 303, a learning unit 304, and a storage unit 310. Specifically, the receiving unit 301, the transmitting unit 302, the identifying unit 303, and the learning unit 304 are realized, for example, by causing the processor 201 to execute a program stored in the storage device 202 shown in FIG. 2, or by the communication IF 205. The storage unit 310 is realized by the storage device 202 shown in FIG. 2.

[0023] The receiving unit 301 receives data from the edge 102 via the network 110, and outputs the data to the identifying unit 303. If the received data is a learning parameter acquisition request or sample information, the receiving unit 301 outputs the data to the identifying unit 303, and if the data is local learning information of a local learning model, the receiving unit 301 outputs the data to the learning unit 304.

[0024] The learning parameter acquisition request includes group information indicating to which attribute group the site ST of the terminal 132, which is the sender, belongs. The group information identifies an attribute group that is, for example, a combination of the industry, use, product type, region, and size of the site ST. The attribute group identifies a learning model in the storage unit 310. The group information may also include a group ID that uniquely identifies the attribute group, rather than the attribute group.

[0025] The local learning information includes the group ID of the site ST and local learning parameters or a local learning data set. The local learning parameters are weights and biases set in the local learning model.

[0026] The sample information includes the group ID of the site ST described above and sample data. The sample data is, for example, a local learning data set used for pre-learning before operation of the local learning model.

[0027] The local learning data set in the local learning information and sample information is a combination of local learning data and local supervised data used for learning the local learning model in the site ST.

[0028] The transmission unit 302 receives the identification result by the identification unit 303 to the edge 102 via the network 110. The identification result is the learning parameters of the learning model identified by the identification unit 303. The learning parameters are weights and biases set in the learning model.

[0029] The identifying unit 303 selects acquisition of learning parameters of a learning model, learning by the learning unit 304, or execution of clustering based on the data from the receiving unit 301. Specifically, for example, if the data from the receiving unit 301 is a learning parameter acquisition request, the identifying unit 303 selects acquisition of learning parameters of a learning model, if the data from the receiving unit 301 is local learning information, the identifying unit 303 selects learning by the learning unit 304, and if the data is sample information, the identifying unit 303 selects execution of clustering.

[0030] When the identification unit 303 selects to acquire the learning parameters of the learning model, the identification unit 303 refers to the group management table 330, identifies the learning model from which the learning parameters are to be acquired from the storage unit 310 using the group information, and acquires the learning parameters of the learning model. The identification unit 303 outputs the acquired learning parameters to the transmission unit 302.

[0031] When the specification unit 303 selects learning by the learning unit 304, it refers to the group management table 330, specifies a learning model to be learned from the storage unit 310 using the group information, and outputs identification information of the specified learning model and the local learning parameters or the local learning data set to the learning unit 304. The identification information of the learning model is a group ID or information corresponding to a group ID.

[0032] When the identification unit 303 selects to perform clustering, it identifies the i-th learning model DB set 310-i corresponding to the group ID included in the sample information. The identification unit 303 clusters the learning data set and sample data of each learning model in the i-th learning model DB set 310-i. By the clustering, a cluster is generated for each attribute group, i.e., each learning model. The identification unit 303 acquires the learning parameters of the learning model corresponding to the cluster to which the sample data belongs.

[0033] The clusters may be generated in advance. In this case, the identification unit 303 acquires the learning parameters of the learning model corresponding to the cluster that is closest to the sample data. The identification unit 303 outputs the acquired learning parameters to the transmission unit 302.

[0034] The learning unit 304 learns the learning model stored in the storage unit 310. Specifically, for example, the learning unit 304 identifies a learning model to be learned by the group ID from the identification unit 303, and learns the learning model by using local learning parameters or a local learning dataset. More specifically, the learning unit 304 executes associative learning in the case of local learning parameters, and executes re-learning (model learning) of the learning model such as regression or discrimination in the case of a local learning dataset.

[0035] The memory unit 310 stores the learning model. Specifically, for example, the memory unit 310 stores the first learning model DB set 310-1 to the n-th learning model DB set 310-n (n is an integer equal to or greater than 1, and satisfies 1≦i≦n. i is an integer equal to or greater than 1). When the first learning model DB set 310-1 to the n-th learning model DB set 310-n are not distinguished from each other, they are referred to as the i-th learning model DB set 310-i. i indicates the type of business (hereinafter, may be referred to as type of business i). In this example, n types of business are applied, and the i-th learning model DB set 310-i exists for each type of business i.

[0036] <i-th learning model DB set 310-i> FIG. 4 is an explanatory diagram showing an example of the ith learning model DB set 310-i. The ith learning model DB set 310-i is a collection of the ith learning model DBs 400-i1 to 400-im (m is an integer equal to or greater than 1, and 1≦j≦m is satisfied. j is an integer equal to or greater than 1). When the ith learning model DBs 400-i1 to 400-im are not distinguished from each other, they are written as the ith learning model DBs 400-ij. j indicates the purpose (hereinafter, may be written as purpose j). In this example, m types of purposes are applied, and the ith learning model DBs 400-ij exist for each purpose j. Note that j may be a different value for each ith learning model DB set 310-i. In other words, the number of types of purpose j may be different for each ith learning model DB set 310-i.

[0037] The i-th learning model DB400-ij has learning model groups 400G-ij1 to 400G-ijk (k is an integer equal to or greater than 1) and a common learning model 403-ij. k indicates variety, region, size, or a combination of these (hereinafter, may be referred to as combination k). Note that k may be a different value for each i-th learning model DB400-ij. In other words, the number of types of combination k may be different for each i-th learning model DB400-ij.

[0038] The learning model group 400G-ijk includes a learning model 401-ijk and a learning dataset 402-ijk. The learning model 401-ijk is a machine learning model trained by the learning dataset 402-ijk. The learning dataset 402-ijk is a combination of learning data and correct answer data. The learning dataset 402-ijk may be data used for training the learning model 401-ijk before the operation of the federated learning system 100, or may be a dataset received from the edge 102 after the operation of the federated learning system 100 has started.

[0039] The common learning model 403-ij is a learning model obtained by averaging the learning parameters of the learning models 401-ij1 to 401-ijk.

[0040] <Group management table 330> 5 is an explanatory diagram showing an example of the group management table 330. The group management table 330 has fields such as group ID 500, business type 501, use 502, product type 503, area 504, and scale 505. For each entry, an attribute group indicating a combination of the business type 501, use 502, product type 503, area 504, and scale 505 is identified.

[0041] A combination of values ​​in the fields 500 to 505 in the same row is an entry that specifies an attribute group that uses one learning model. Note that in each entry, there may be cases where no value exists in any of the business type 501, use 502, product type 503, region 504, and size 505.

[0042] The group ID 500 is identification information that uniquely identifies an attribute group, that is, a learning model in the storage unit 310. Therefore, the server 101 can identify a learning model in the storage unit 310 by using the group ID 500 as a key. For convenience of explanation, each value of the group ID 500 is a code of a corresponding learning model.

[0043] <Learning Section 304> 6 is a block diagram showing an example of a functional configuration of the learning unit 304. The learning unit 304 has a determination unit 600, an associative learning unit 601, and an individual learning unit 602. The determination unit 600 determines whether the data from the receiving unit 301 is a local learning parameter or a local learning data set. If the data is a local learning parameter, the determination unit 600 outputs the local learning parameter to the associative learning unit 601, and if the data is a local learning data set, the determination unit 600 outputs the local learning parameter to the individual learning unit 602.

[0044] The associative learning unit 601 identifies a learning model to be subjected to associative learning from the group ID 500, and executes associative learning of the identified learning model based on the local learning parameters. Specifically, for example, the associative learning unit 601 adds a group of local learning parameters used so far by the learning model to be subjected to associative learning to the local learning parameters given this time, and divides the result by the number of local learning parameters after the addition to calculate an averaged local learning parameter. The associative learning unit 601 updates the learning model to be subjected to associative learning with the averaged local learning parameters.

[0045] The individual learning unit 602 identifies a learning model to be learned from the group information, and executes federated learning of the identified learning model based on the local learning data set. Specifically, for example, the individual learning unit 602 inputs local learning data in the local learning data set to the learning model to be learned, and calculates the value of a loss function based on the difference between the learning result output from the learning model to be learned and the correct answer data in the local learning data set. The individual learning unit 602 updates the learning model to be learned by executing error backpropagation so as to reduce the value of the loss function.

[0046] In addition, when the above-mentioned updates are performed by either the associative learning unit 601 or the individual learning unit 602, the associative learning unit 601 updates the common learning model 403-ij by averaging the learning parameters of the learning models 401-ij1 to 401-ijk in the i-th learning model DB 400-ij to which the learning model to be learned by the associative learning belongs.

[0047] <Edge 102> 7 is a block diagram showing an example of a functional configuration of the edge 102-f. The edge 102-f has a local prediction unit 701, a local learning unit 702, a local management unit 703, an analysis DB 704, and a local learning model DB 705. Specifically, the local prediction unit 701, the local learning unit 702, and the local management unit 703 are realized by, for example, making the processor 201 execute a program stored in the storage device 202 shown in FIG. 2. The analysis DB 704 and the local learning model DB 705 are realized by the storage device 202 shown in FIG. 2.

[0048] The local prediction unit 701 obtains the local learning model 750 specified by the terminal 132-f from the local learning model DB 705, and calculates a prediction result by inputting data from the sensor 131-f to the local learning model 750. The prediction result is stored in the analysis DB 704.

[0049] The local learning unit 702 acquires a local learning model corresponding to the sensor 131-f from the local learning model DB 705, and calculates a value of a loss function based on a difference between the data from the sensor 131-f and a prediction result. The local learning unit 702 updates the local learning model to be learned by executing error backpropagation so as to reduce the value of the loss function.

[0050] The local management unit 703 acquires local learning parameters of the local learning model to be the subject of associative learning from the local learning model DB 705. The local management unit 703 transmits to the server 101 a learning parameter acquisition request including the group ID 500, which is the group information of the site STf, and the acquired local learning parameters.

[0051] In addition, the local management unit 703 acquires a local learning data set of the local learning model to be subjected to the associative learning from the local learning model DB 705. The local management unit 703 transmits to the server 101 local learning information including the group ID 500, which is group information of the site STf, and the acquired local learning data set.

[0052] In addition, there are cases where the administrator of the site STf does not know which attribute group the local learning model 750 belongs to. In this case, the local management unit 703 transmits sample information to the server 101 in response to an instruction from the terminal 132-f.

[0053] The analysis DB 704 stores the prediction result of the local prediction unit 701. The analysis DB 704 also stores the data from the sensor 131-F input to the local prediction unit 701 in association with the prediction result. The analysis DB 704 also stores the correct answer data of the data from the sensor 131-F from the terminal 132-f in association with the data from the sensor 131-F. The data from the sensor 131-F becomes local learning data, and the combination of the data from the sensor 131-F and the correct answer data becomes a local learning data set.

[0054] The local learning model DB 705 stores local learning models 750 learned by the local learning unit 702. The number of local learning models exists equal to the number of applications 502 in the site STf.

[0055] <First learning sequence> 8 is a sequence diagram showing an example of a first learning sequence in the federated learning system 100. The first learning sequence is a sequence showing learning based on a learning parameter acquisition request.

[0056] The edge 102-f transmits a learning parameter acquisition request to the server 101 (step S801).

[0057] The server 101 receives the learning parameter acquisition request via the receiving unit 301 (step S802).

[0058] The server 101 selects, by the specification unit 303, acquisition of the learning parameters of the learning model (step S803).

[0059] If the group information in the learning parameter acquisition request includes the group ID 500 (step S804: Yes), the server 101 proceeds to step S806. If the group information in the learning parameter acquisition request does not include the group ID 500 (step S804: No), the server 101 identifies the group ID 500 corresponding to the attribute group from the group management table 330 (step S805), and proceeds to step S806.

[0060] Note that multiple group IDs 500 corresponding to an attribute group may be identified. For example, if the attribute group is (industry W1, use U1, product type V1), it is unclear whether the group corresponds to area Z1 to Zc or scale S1 to Sd. In this case, a group ID 500 that satisfies (industry W1, use U1, product type V1, area Z1 to Zc, scale S1 to Sd) is identified.

[0061] The server 101, by the specification unit 303, acquires the learning parameters of the learning model corresponding to the group ID 500 from the storage unit 310 (step S806). In this case, the server 101 also acquires the group ID 500 and group information that is its attribute group.

[0062] The server 101 transmits the acquired group information and learning parameters to the edge 102-f that has transmitted the learning parameter acquisition request, via the transmission unit 302 (step S807).

[0063] The edge 102-f receives the group information and the learning parameters (step S808). The edge 102-f updates a local learning model corresponding to the usage of the attribute group with the learning parameters (step S809).

[0064] If multiple learning parameters are acquired, the edge 102-f transmits multiple attribute groups corresponding to the acquired learning parameters to the terminal 132-f prior to step S809. The terminal 132-f displays the multiple attribute groups and waits for a selection from a user of the terminal 132-f (e.g., an administrator of the site STf). The terminal 132-f transmits the selected attribute group to the edge 102-f. The edge 102-f updates a local learning model corresponding to the use of the selected attribute group with the learning parameters corresponding to the selected attribute group.

[0065] <Second learning sequence> 9 is a sequence diagram showing an example of a second learning sequence in the federated learning system 100. The second learning sequence is a sequence showing learning based on local learning information. The edge 102-f transmits the local learning information to the server 101 (step S901).

[0066] The server 101 receives the local learning information through the receiving unit 301 (step S902).

[0067] The server 101 selects learning by the learning unit 304 through the identifying unit 303 (step S903).

[0068] If the group information of the local learning information includes the group ID 500 (step S904: Yes), the server 101 proceeds to step S906. If the group information of the local learning information does not include the group ID 500 (step S904: No), the server 101 identifies the group ID 500 corresponding to the attribute group from the group management table 330 (step S905), and proceeds to step S906.

[0069] The case where a plurality of group IDs 500 corresponding to the attribute group are identified is similar to the case shown in step S805 of FIG.

[0070] The server 101 uses the learning unit 304 to learn the learning model corresponding to the group ID 500 using the local learning information (step S906). Specifically, for example, if the local learning information is a local learning parameter, the server 101 uses the associative learning unit 601 to update the learning model to be subjected to associative learning corresponding to the group ID 500, as described in FIG.

[0071] If the local learning information is a local learning parameter, the server 101 updates the learning model of the learning target corresponding to the group ID 500 by the individual learning unit 602 as described in Fig. 6. Also, the server 101 updates the common learning model 403-ij with learning parameters obtained by averaging the learning parameter group of the updated learning model group 402-ij1 to 402-ijk.

[0072] The server 101, by the specification unit 303, acquires the learning parameters of the learning model corresponding to the group ID 500 from the storage unit 310 (step S806). In this case, the server 101 also acquires the group ID 500 and group information that is its attribute group.

[0073] Thereafter, the server 101 executes steps S806 and S807 for the learning parameters of the updated learning model, and the edge 102-f executes steps S808 and S809.

[0074] <Third Learning Sequence> 10 is a sequence diagram showing an example of a third learning sequence in the federated learning system 100. The third learning sequence is a sequence showing learning based on sample information. The edge 102-f transmits the sample information to the server 101 (step S1001).

[0075] The server 101 receives the sample information through the receiving unit 301 (step S1002).

[0076] The server 101 selects execution of clustering by the specifying unit 303 (step S1003).

[0077] If the group information of the sample information includes the group ID 500 (step S1004: Yes), the server 101 proceeds to step S1006. If the group information of the sample information does not include the group ID 500 (step S1004: No), the server 101 identifies the group ID 500 corresponding to the attribute group from the group management table 330 (step S1005), and proceeds to step S1006.

[0078] The case where a plurality of group IDs 500 corresponding to the attribute group are identified is similar to the case shown in step S805 of FIG.

[0079] The server 101 causes the identification unit 303 to identify the i-th learning model DB 400-ij corresponding to the group ID 500 from the storage unit 310 (step S1006).

[0080] The server 101 executes clustering based on the learning data set group in the i-th learning model DB 400-ij and the sample data by the identification unit 303 (step S1007). Specifically, for example, the server 101 identifies a learning model corresponding to a cluster included in the sample data from the i-th learning model DB 400-ij. Then, the server 101 acquires the learning parameters of the identified learning model and the group ID 500 corresponding to the identified learning model (step S1007).

[0081] The cluster group may be generated in step S1007 including the sample data, or may be generated in advance. When the cluster group is generated in advance, a learning model corresponding to a cluster that is closest to the sample data in the feature space is identified.

[0082] Thereafter, the server 101 executes steps S806 and S807 for the learning parameters of the learning model identified by clustering, and the edge 102-f executes steps S808 and S809.

[0083] In the third learning sequence, the server 101 may transmit only the group information out of the group information and the learning parameters in step S807.

[0084] Also, the edge 102-f may transmit the attribute group excluding the sample data, although the sample information was transmitted in step S1001. As a result, the server 101 identifies the group ID 500 corresponding to the attribute group (step S1005), and transmits the group ID 500 to the edge 102-f (step S807) without executing steps S1006 and S1007.

[0085] This allows the user of the terminal 132-f at the site STf to identify which group ID corresponds to the attribute group transmitted to the server 101, and the terminal 132-f can set the group ID 500 in the edge 102-f. Therefore, from the next time onwards, the edge 102-f can transmit a learning parameter acquisition request or local learning information to the server 101 by using the group ID 500.

[0086] Furthermore, if multiple group IDs 500 are identified in step S1005, the server 101 also transmits the attribute group (industry 501 to scale 505) for each group ID 500 to the edge 102-f. This allows the user of the terminal 132-f of the site STf to check which attribute group (industry 501 to scale 505) in the group management table 330 the attribute group transmitted to the server 101 corresponds to for each group ID, and select an appropriate attribute group. Then, the terminal 132-f can set the group ID 500 of the attribute group selected by the user to the edge 102-f. Therefore, from the next time onwards, the edge 102-f can transmit a learning parameter acquisition request or local learning information to the server 101 using the group ID 500.

[0087] 8 to 10, the server 101 transmits the learning parameters to the edge 102-f, but the server 101 may transmit the learning model in which the learning parameters are set to the edge 102-f. In this case, the edge 102-f sets the received learning model as the local learning model.

[0088] Also, in Figures 8 to 10, for a user who only knows the attribute group as industry 501, when the attribute group received from edge 102-f is only industry 501, the server 101 may send the learning parameters of the common learning model 403-ij in the i-th learning model DB set 310-i corresponding to industry 501 to edge 102-f.

[0089] As described above, according to this embodiment, the server 101 has an ith learning model DB set 310-i for each industry that has been trained in advance using data from various sites. This allows the local learning model to be trained using a learning model suitable for the attribute group (industry 501 to size 505) of the site STf.

[0090] Furthermore, by deploying a local learning model in the edge 102-f, the site STf can update the local learning model with the learning model of the server 101 without transferring data outside the site STf. Therefore, the learning model of the server 101, which has a high accuracy rate, can be utilized without worrying about data leaking outside the site STf. Also, since data is not transferred outside the site STf, it is possible to reduce the amount of communication.

[0091] In addition, since the learning models of the server 101 are associated with the group ID 500, the user can determine which learning model is suitable for the site STf.

[0092] In this way, regardless of whether a user of site STf may send learning parameters of a local learning model to server 101, a user of site STf may send a learning data set to server 101, or a user of site STf may send sample data to server 101, server 101 can obtain appropriate learning parameters and provide them to the user.

[0093] The present invention is not limited to the above-described embodiments, and includes various modified examples and equivalent configurations within the spirit of the appended claims. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to those having all of the configurations described. Also, a part of the configuration of one embodiment may be replaced with a configuration of another embodiment. Also, a configuration of another embodiment may be added to a configuration of one embodiment. Also, a part of the configuration of each embodiment may be added, deleted, or replaced with another configuration.

[0094] Furthermore, each of the aforementioned configurations, functions, processing units, processing means, etc. may be realized in hardware, for example by designing some or all of them as an integrated circuit, or may be realized in software by a processor interpreting and executing a program that realizes each function.

[0095] Information such as programs, tables, files, etc. that realize each function can be stored in a storage device such as a memory, a hard disk, or an SSD (Solid State Drive), or in a recording medium such as an IC (Integrated Circuit) card, an SD card, or a DVD (Digital Versatile Disc).

[0096] In addition, the control lines and information lines shown are those considered necessary for the explanation, and do not necessarily show all the control lines and information lines necessary for implementation. In reality, it can be considered that almost all components are connected to each other. [Explanation of symbols]

[0097] STf site 100 Federated Learning System 101 Server (management device) 102-f Edge 131-f Sensor 132-f terminal 301 Receiving unit 302 Transmitter 303 Specific section 304 Learning Department 310 Storage section 310-i ith learning model DB set 330 Group Management Table 400-ij ith learning model DB 402-ij1~402-ijk Learning model 500 Group ID 600 Judgment section 601 Joint Learning Department 602 Individual Study Department 701 Local Prediction Department 702 Local Learning Department 703 Local Management Department 704 Analysis DB 705 Local Learning Model DB 750 Local Learning Model

Claims

1. A storage unit that stores a plurality of first learning models trained for each group ID that identifies an attribute group that is a combination of a plurality of attributes; A receiving unit that receives group information including the group ID or an attribute group; an identification unit that identifies a first learning parameter of the first learning model corresponding to the group ID based on the group information received by the receiving unit; a transmission unit that transmits the first learning parameter identified by the identification unit to a source of the group information; A management device comprising:

2. The management device according to claim 1 , The group information includes the group ID, The identification unit identifies a first learning parameter of the first learning model corresponding to the group ID received by the receiving unit. A management device comprising:

3. The management device according to claim 2, A learning unit that learns the first learning model, The source has a second learning model; The receiving unit receives a second learning parameter of the second learning model from the transmission source; The learning unit learns the first learning model based on the second learning parameters and the first learning parameters; The transmission unit transmits a first learning parameter of the first learning model after learning by the learning unit to the transmission source. A management device comprising:

4. The management device according to claim 2, A learning unit that learns the first learning model trained with a first learning data set that is a combination of first learning data and first correct answer data, The source has a second learning model; The receiving unit receives, from the transmission source, a second learning data set which is a combination of second learning data and second supervised data used in learning the second learning model; The learning unit learns the first learning model based on the first learning data set and the second learning data set; The transmission unit transmits a first learning parameter of the first learning model after learning by the learning unit to the transmission source. A management device comprising:

5. The management device according to claim 1 , the storage unit stores group management information that associates the group ID with the attribute group; the group information includes the attribute group; The identification unit refers to the group management information to identify a group ID associated with the attribute group received by the reception unit; The transmission unit transmits the group ID acquired by the identification unit to the transmission source. A management device comprising:

6. The management device according to claim 5 , The receiving unit receives an attribute group including fewer attributes than the plurality of attributes, The identification unit refers to the group management information to identify a plurality of attribute groups including the attribute group and a plurality of group IDs associated with each of the plurality of attribute groups; The transmission unit transmits the plurality of attribute groups and the plurality of group IDs to the transmission source. A management device comprising:

7. The management device according to claim 6, The identification unit identifies a first learning parameter of the first learning model corresponding to each of a plurality of group IDs; The transmission unit transmits the plurality of first learning parameters acquired by the identification unit to the transmission source. A management device comprising:

8. The management device according to claim 5 , A learning unit that learns the first learning model, The source has a second learning model; The receiving unit receives a second learning parameter of the second learning model from the transmission source; The learning unit learns the first learning model based on the second learning parameters and the first learning parameters; The transmission unit transmits a first learning parameter of the first learning model after learning by the learning unit to the transmission source. A management device comprising:

9. The management device according to claim 5 , A learning unit that learns the first learning model trained with a first learning data set that is a combination of first learning data and first correct answer data, The source has a second learning model; The receiving unit receives, from the transmission source, a second learning data set which is a combination of second learning data and second supervised data used in learning the second learning model; The learning unit learns the first learning model based on the first learning data set and the second learning data set; The transmission unit transmits a first learning parameter of the first learning model after learning by the learning unit to the transmission source. A management device comprising:

10. The management device according to claim 8, the storage unit stores the plurality of first learning models for each specific attribute among the plurality of attributes, and stores a common first learning model trained by a plurality of first learning parameters of the plurality of first learning models for each specific attribute; The receiving unit receives the specific attribute as the attribute group, The identification unit identifies the common first learning model corresponding to the specific attribute, The transmission unit transmits learning parameters of the common first learning model to the transmission source. A management device comprising:

11. A management device having a storage unit that stores a plurality of first learning models trained for each group ID that identifies an attribute group that is a combination of a plurality of attributes, A receiving process of receiving group information including the group ID or attribute group; A process of identifying a first learning parameter of the first learning model corresponding to the group ID based on the group information received by the receiving process; a transmission process of transmitting the first learning parameter identified by the identification process to a source of the group information; A management method comprising the steps of:

12. A processor that can access a storage unit that stores a plurality of first learning models trained for each group ID that identifies an attribute group that is a combination of a plurality of attributes, A receiving process of receiving group information including the group ID or attribute group; A process of identifying a first learning parameter of the first learning model corresponding to the group ID based on the group information received by the receiving process; a transmission process of transmitting the first learning parameter identified by the identification process to a source of the group information; A management program for causing the user to execute the above steps.

Images