Terminal device, federated learning system, server, and model update method

Abstract

This terminal device (100) is provided with: a global model reception unit (110) that receives a global model updated by integrating a plurality of learned local models, each of which is a local model learned at a different terminal device; a global model evaluation unit (120) that evaluates the global model received by the global model reception unit and a local model, which is currently retained, to determine whether or not to adopt the global model; and a local model update unit (130) that determines an update method of a local model on the basis of an evaluation result by the global model evaluation unit, and updates the local model being retained, according to the update method being determined.

Description

Terminal device, federated learning system, server, and model update method

[0001] The disclosed technology relates to a federated learning technology that updates a learning model using a global model constructed based on multiple local models.

[0002] Federated learning technology is used, for example, when it is difficult to collect sufficient training data, or when data collected by multiple different organizations cannot be aggregated and used as training data due to considerations such as personal information protection. In federated learning, a common model is generally used as an initial value, and a global model on a server is updated based on local models trained on multiple clients. The updated global model is then distributed to the clients and adopted as their local models. Meanwhile, Patent Document 1 ("Analysis Device, Analysis Method, and Analysis Program") discloses a generation method for generating an appropriate predictive model according to the characteristics of training data D from each location S before federated learning. Specifically, in the generation method of Patent Document 1, distribution analysis is performed on the server based on the transformation features of each training data from each location, and a learning method for each location is predetermined according to the distribution similarity.

[0003] Japanese Patent Application Laid-Open No. 2023-173559 (Hitachi, Ltd.)

[0004] However, conventional techniques have a problem in that, for example, if the data distribution between clients differs over time and becomes non-independent and identical, the accuracy of the learning model for the clients decreases. The generation method of Patent Document 1 cannot solve the above problem because the learning method is determined in advance before federated learning.

[0005] The present disclosure is intended to solve the above-mentioned problems and aims to improve the accuracy of learning models for clients compared to conventional federated learning.

[0006] The terminal device of the present disclosure includes: a global model reception unit that receives an updated global model by integrating multiple trained local models, each of which is a local model trained in a different terminal device; a global model evaluation unit that determines whether to adopt the global model based on evaluation results of the global model received by the global model reception unit and a currently held local model; and a local model update unit that determines a method for updating the local model based on the evaluation results by the global model evaluation unit, and updates the held local model in accordance with the determined update method.

[0007] The present disclosure has the effect of improving the accuracy of the learning model for the client compared to conventional federated learning.

[0008] FIG. 1 is a diagram illustrating an example of the configuration of a terminal device according to a first embodiment of the present disclosure. FIG. 2 is a flowchart illustrating an example of a process for realizing characteristic functions in the terminal device according to the first embodiment of the present disclosure. FIG. 3 is a diagram illustrating an example of the configuration of a federated learning system including the terminal device according to the first embodiment of the present disclosure. FIG. 4 is a diagram illustrating an example of the internal configuration of the terminal device and an example of the internal configuration of the server according to the first embodiment of the present disclosure. FIG. 5 is a flowchart illustrating an example of a process in the terminal device according to the first embodiment of the present disclosure. FIG. 6 is a flowchart illustrating a detailed example of a process from a global evaluation process to a local model update process in the terminal device according to the first embodiment of the present disclosure. FIG. 7 is a sequence diagram illustrating an example of a process in a federated learning system including a terminal device and a server according to the first embodiment of the present disclosure. FIG. 8 is a diagram illustrating an example of the configuration of a federated learning system including a terminal device according to the second embodiment of the present disclosure. FIG. 9 is a diagram for explaining a model update method by the terminal device according to the second embodiment of the present disclosure. FIG. 10 is a flowchart illustrating an example of an individualization update process in the terminal device according to the second embodiment of the present disclosure. FIG. 11 is a flowchart illustrating an example of a process in a federated learning system including a terminal device and a server according to the second embodiment of the present disclosure. FIG. 12 is a diagram illustrating an example of the configuration of a federated learning system including a terminal device according to a third embodiment of the present disclosure. FIG. 13 is a flowchart illustrating an example of processing in a federated learning system including a terminal device and a server according to the third embodiment of the present disclosure. FIG. 14 is a sequence diagram illustrating an example of processing in a federated learning system including a terminal device and a server according to the third embodiment of the present disclosure. FIG. 15 is a diagram illustrating an example of the configuration of a federated learning system including a terminal device according to a fourth embodiment of the present disclosure. FIG. 16 is a diagram for explaining a model update method by a terminal device according to the fourth embodiment of the present disclosure. FIG. 17 is a flowchart illustrating a detailed example of processing from global evaluation processing to local model update processing in a terminal device according to the fourth embodiment of the present disclosure. FIG. 18 is a flowchart illustrating an example of individualization update processing in a terminal device according to the fourth embodiment of the present disclosure. FIG. 19 is a diagram illustrating a first example of a hardware configuration for realizing functions according to the configuration of the present disclosure.FIG. 20 is a diagram illustrating a second example of a hardware configuration for realizing the functions according to the configuration of the present disclosure.

[0009] In order to explain the present disclosure in more detail, embodiments of the present disclosure will be described below with reference to the accompanying drawings.

[0010] Embodiment 1. In embodiment 1, a configuration example of a basic form for realizing the technology of the present disclosure will be described. A configuration example of an apparatus according to embodiment 1 of the present disclosure will be described. FIG. 1 is a diagram showing an example of the configuration of a terminal apparatus 100 according to embodiment 1 of the present disclosure. The terminal apparatus 100 shown in FIG. 1 is configured to include a global model receiving unit 110, a global model evaluating unit 120, a local model updating unit 130, a local model holding unit 140, a data collecting unit 150, a local model learning unit 160, and a local model information output unit 170.

[0011] The global model receiving unit 110 has a function of receiving a global model. The global model receiving unit 110 receives a global model that has been updated by integrating multiple trained local models, each of which is a local model trained in a different terminal device (100).

[0012] The global model evaluation unit 120 has a function of evaluating a global model. The global model evaluation unit 120 evaluates the global model received by the global model receiving unit 110 and the currently held local model, and determines whether to adopt the global model based on the evaluation results. The evaluation results are output in the form of, for example, an evaluation value. The model evaluation method may be any known model evaluation technique, as long as it outputs an evaluation result that indicates whether the performance of the global model is lower or higher than the performance of the currently held local model.

[0013] The local model update unit 130 determines a method for updating the local model and updates the local model. The local model update unit 130 determines a method for updating the local model based on the evaluation result by the global model evaluation unit 120, and updates the currently held local model according to the determined update method. The update method may be, for example, full adoption (normal federated learning), partial adoption (individualized federated learning), or full rejection (client-only learning).

[0014] The local model holding unit 140 holds the local model and the local model updated by the local model updating unit 130.

[0015] The data collection unit 150 collects learning data for training a local model. The data collection unit 150 collects learning data for training a local model held in the local model holding unit 140. The data collection unit 150 also collects evaluation data used when the global model evaluation unit 120 evaluates a global model. The evaluation data is, for example, a data set that combines input data to be input to a model and output data that is data to be output.

[0016] The local model learning unit 160 has a function of learning a local model. The local model learning unit 160 learns and constructs the local model held in the local model holding unit 140 based on the learning data collected by the data collecting unit 150.

[0017] The local model information output unit 170 outputs a trained local model, which is a local model trained on the client side. The local model information output unit 170 outputs the trained local model, which is a local model trained by the local model training unit 160, to the outside of the device.

[0018] In addition to the above components, the terminal device 100 also includes a terminal-side control unit (not shown), a terminal-side storage unit (not shown), and a terminal-side communication unit (not shown). The terminal-side control unit (not shown) controls the entire terminal device 100 and each of its components. The terminal-side control unit (not shown), for example, starts up the terminal device 100 in accordance with an external command. The terminal-side control unit (not shown) also controls the state of the terminal device 100 (operating state = startup, shutdown, sleep, etc.). The terminal-side storage unit (not shown) stores various data used by the terminal device 100. The terminal-side storage unit (not shown), for example, stores output (output data) from each component of the terminal device 100 and outputs data requested for transmission or acquisition by each component to the component corresponding to the request. The terminal-side communication unit (not shown) communicates with external devices. For example, it communicates between the terminal device 100 and peripheral devices (e.g., input / output devices (not shown)). For example, when the terminal device 100 and the input / output device are not connected by wire, the terminal-side communication unit (not shown) has a function of communicating between the terminal device 100 and the input / output device. Also, when the terminal device 100 and a server (server device, cloud server) cooperate to realize the federated learning technology of the present disclosure, the terminal-side communication unit (not shown) has a function of communicating between the terminal device 100 and the server. The terminal-side control unit (not shown), the terminal-side storage unit (not shown), and the terminal-side communication unit (not shown) each have the same functions in the embodiments described below.

[0019] Next, an example of a process for realizing a characteristic function in the terminal device according to the first embodiment of the present disclosure will be described. FIG. 2 is a flowchart illustrating an example of a process for realizing a characteristic function in the terminal device according to the first embodiment of the present disclosure. The process illustrated in FIG. 2 is a model update method for updating a local model used in the terminal device 100 based on a global model received from outside the device. The terminal device 100 starts the process illustrated in FIG. 2 ("Start"), for example, when it receives a command operation from a user or receives a global model from outside the device and the global model receiving unit 110 receives the global model. The global model receiving unit 110 receives an updated global model by integrating multiple trained local models, each of which is a local model trained in a different terminal device (100). The global model receiving unit 110 outputs the global model to the global model evaluation unit 120.

[0020] The terminal device 100 then executes a global model evaluation process (step ST100). In the global model evaluation process, the global model evaluation unit 120 of the terminal device 100 evaluates the global model accepted by the global model acceptance unit 110 and the currently held local model, and determines whether to accept the global model based on the evaluation results. The global model evaluation unit 120 outputs the evaluation results to the local model update unit 130.

[0021] The terminal device 100 then executes an update method determination process according to the evaluation result (step ST200). In the update method determination process, the local model update unit 130 of the terminal device 100 determines an update method for the local model according to the evaluation result by the global model evaluation unit 120. If the evaluation result by the global model evaluation unit 120 indicates that the performance of the global model is higher than that of the local model, the local model update unit 130 adopts an update method of replacing the local model with the global model. If the evaluation result by the global model evaluation unit 120 indicates that the performance of the global model is lower than that of the local model, the local model update unit 130 does not adopt an update method of directly replacing the local model with the global model. The local model update unit 130 determines an update method such as full adoption (normal federated learning), partial adoption (individualized federated learning), or full rejection (client-only learning).

[0022] The terminal device 100 then executes a local model update process (step ST300). In the local model update process, the local model update unit 130 of the terminal device 100 updates the currently held local model in accordance with the determined update method. The updated local model is held in the local model holding unit 140.

[0023] After executing the local model update process, the terminal device 100 then ends the process shown in FIG. 2 ("End").

[0024] Next, an example of the configuration of a federated learning system including a terminal device according to the first embodiment of the present disclosure will be described. FIG. 3 is a diagram illustrating an example of the configuration of a federated learning system 1 (1A) including a terminal device 100 (100A) according to the first embodiment of the present disclosure. The federated learning system 1 (1A) is a system that performs federated learning using a local model on the terminal device 100 (100A) side and a global model on the server 300 side. The federated learning system 1 (1A) collects local models on the terminal device 100 (100A) side, updates a global model using the local models, and distributes the global model to the terminal device 100 (100A). The federated learning system 1 (1A) illustrated in FIG. 3 illustrates an example of a provider-side configuration and a client-side configuration for federated learning. The provider-side configuration is used by a provider that distributes and provides a global model to multiple clients. The client-side configuration is used by a client that accepts a global model. In the federated learning system 1 (1A) shown in FIG. 3, the client side includes a terminal device 100 and an information source device 200 for each client. For example, client 1 includes a terminal device 100A-1 and an information source device 200-1, client 2 includes a terminal device 100A-2 and an information source device 200-2, client 3 includes a terminal device 100A-3 and an information source device 200-3, ..., and client n includes a terminal device 100A-n and an information source device 200-n. In the federated learning system 1 (1A), the provider side includes a server 300. The terminal device 100 and the information source device 200 are communicatively connected. Furthermore, the terminal devices 100 (100A-1, 100A-2, 100A-3, ..., 100A-n (n≧1)) and the server 300 are communicatively connected. Since the multiple terminal devices 100 (100A-1, 100A-2, 100A-3, . . . , 100A-n (n≧2)) each have the same configuration, the following description will be given with one terminal device 100 and one server 300 shown.

[0025] 4 is a diagram illustrating an example of the internal configuration of a terminal device and an example of the internal configuration of a server according to the first embodiment of the present disclosure. The federated learning system 1 (1A) illustrated in FIG. 4 includes a terminal device 100 (100A), an information source device 200, and a server 300.

[0026] The terminal device 100 (100A) is configured to have the same functions as the already-described terminal device 100. The terminal device 100 (100A) shown in Fig. 4 is configured to include a global model receiving unit 110, a global model evaluating unit 120, a local model updating unit 130, a local model holding unit 140, a data collecting unit 150, a local model learning unit 160, and a local model information output unit 170.

[0027] The global model receiving unit 110 receives a global model that has been updated by integrating a plurality of local models that have been trained in different terminal devices (100A-1, . . . , 100A-n (n≧1)).

[0028] The global model evaluation unit 120 evaluates the global model received by the global model reception unit 110 and the currently held local model, and determines whether to adopt the global model based on the evaluation result. The evaluation result is output in the form of, for example, an evaluation value.

[0029] The local model update unit 130 determines a method for updating the local model based on the evaluation result by the global model evaluation unit 120, and updates the held local model in accordance with the determined update method.

[0030] The local model holding unit 140 holds the local model updated by the local model updating unit 130 .

[0031] The data collection unit 150 collects, from the information source device 200, learning data for training the local model held in the local model holding unit 140.

[0032] The local model learning unit 160 builds the local model held in the local model holding unit 140 by learning it based on the learning data collected by the data collecting unit 150 .

[0033] The local model information output unit 170 outputs the trained local model, which is the local model trained by the local model training unit 160, to the server 300 outside the device.

[0034] The information source device 200 outputs data used in the terminal device 100 (100A). Specifically, the information source device 200 outputs data including training data for training a local model in the terminal device 100 (100A). The information source device 200 also outputs a dataset that combines training data and teacher data used to evaluate the global model and the local model. The federated learning system 1 (1A) has an information source device 200 for each client, and is equipped with multiple information source devices 200-1, 200-2, 200-3, ..., 200-n. The information source device 200 in the federated learning system 1 (1A) is, for example, a surveillance camera that captures images of a monitored area. In the case of a surveillance camera, the training data is a captured image, and the dataset is a combination of the captured image and information indicating an area requiring attention.

[0035] The server 300 receives local models from the client terminal devices 100 (100A), integrates multiple local models to update the global model, and distributes the updated global model to each of the terminal devices 100 (100A). The server 300 shown in FIG. 4 includes a local model information receiving unit 310, a global model integration unit 320, and a global model distribution unit 330.

[0036] The local model information receiving unit 310 receives the trained local model trained in the terminal device 100 (100A).

[0037] The global model integration unit 320 integrates the trained local models trained in the different terminal devices (100A-1, . . . , 100A-n (n≧2)) to update the global model.

[0038] The global model distribution unit 330 outputs the global model updated by the global model integration unit 320 to the terminal devices 100 (100A-1, . . . , 100A-n (n≧1)).

[0039] In addition to the above components, the server 300 also includes a server-side control unit (not shown), a server-side storage unit (not shown), and a server-side communication unit (not shown). The server-side control unit (not shown) controls the entire server 300 and each of its components. The server-side control unit (not shown), for example, starts up the server 300 in accordance with external commands. The server-side control unit (not shown) also controls the state of the server 300 (operating state, such as startup, shutdown, or sleep). The server-side storage unit (not shown) stores data used by the server 300. The server-side storage unit (not shown), for example, stores output (output data) from each component of the server 300 and outputs data requested for sending or retrieval by each component to the component corresponding to the request. The server-side communication unit (not shown) communicates with external devices. For example, the server 300 communicates with a peripheral device (e.g., an input / output device (not shown). For example, if the server 300 and the input / output device are not connected via a wired connection, the server-side communication unit (not shown) has the function of communicating between the server 300 and the input / output device. Furthermore, when the server 300 (server device, cloud server) and the terminal device 100 (100A) cooperate to realize the federated learning technology of the present disclosure, a server-side communication unit (not shown) has a function of communicating between the server 300 and the terminal device 100 (100A). The server-side control unit (not shown), the server-side storage unit (not shown), and the server-side communication unit (not shown) each have the same functions in the embodiments described below.

[0040] Next, an example of processing of the terminal device according to the first embodiment of the present disclosure will be described. FIG. 5 is a flowchart showing an example of processing of the terminal device according to the first embodiment of the present disclosure. The processing shown in FIG. 5 is a model update method for updating a local model used in the terminal device 100 (100A) based on a global model received in the terminal device 100 (100A) from outside the device (server 300). The terminal device 100 (100A) starts the processing shown in FIG. 5 ("Start"), for example, in response to a user command operation or reception of a global model from outside the device.

[0041] The terminal device 100 (100A) then executes a global model reception process (step ST1100). In the global model reception process, the global model reception unit 110 of the terminal device 100 (100A) receives an updated global model by integrating multiple trained local models, each of which is a local model trained in a different terminal device (100A-1, ..., 100A-n). The global model reception unit 110 outputs the global model to the global model evaluation unit 120.

[0042] The terminal device 100 (100A) then executes a global model evaluation process (step ST1110). In the global model evaluation process, the global model evaluation unit 120 of the terminal device 100 (100A) evaluates the global model accepted by the global model acceptance unit 110 and the currently held local model, and determines whether to accept the global model based on the evaluation results. The global model evaluation unit 120 outputs the evaluation results to the local model update unit 130.

[0043] The terminal device 100 (100A) then executes an update method determination process according to the evaluation result (step ST1120). In the update method determination process according to the evaluation result, the local model update unit 130 of the terminal device 100 (100A) determines an update method for the local model based on the evaluation result by the global model evaluation unit 120. If the evaluation result by the global model evaluation unit 120 indicates that the performance of the global model is higher than that of the local model, the local model update unit 130 adopts an update method that replaces the local model with the global model. If the evaluation result by the global model evaluation unit 120 indicates that the performance of the global model is lower than that of the local model, the local model update unit 130 does not adopt an update method that directly replaces the local model with the global model. The local model update unit 130 determines an update method such as full adoption (normal federated learning), partial adoption (individualized federated learning), or full rejection (client-only learning), and updates the held local model according to the determined update method.

[0044] The terminal device 100 (100A) then executes a local model update process (step ST1130). In the local model update process, the local model update unit 130 of the terminal device 100 (100A) updates the currently held local model in accordance with the determined update method. The updated local model is held in the local model holding unit 140.

[0045] The terminal device 100 (100A) then executes a data collection process (step ST1140). In the data collection process, the data collection unit 150 of the terminal device 100 (100A) collects, from the information source device 200, learning data for training the local model held in the local model holding unit 140.

[0046] The terminal device 100 (100A) then executes a local model learning process (step ST1150). In the local model learning process, the local model learning unit 160 of the terminal device 100 (100A) learns and constructs the local model held in the local model holding unit 140 based on the learning data collected by the data collecting unit 150.

[0047] The terminal device 100 (100A) then executes a local model information output process (step ST1160). In the local model information output process, the local model information output unit 170 of the terminal device 100 (100A) outputs a trained local model, which is the local model trained by the local model training unit 160, to the server 300 outside the device.

[0048] The terminal device 100 (100A) then executes an end determination process (step ST1170 "End?"). In the end determination process, a terminal-side control unit (not shown) of the terminal device 100 (100A) determines whether to end the processing of the terminal device 100 (100A). The terminal-side control unit (not shown) determines whether to end the processing of the terminal device A, for example, in accordance with an external end command or an execution program. If the terminal-side control unit (not shown) determines that the processing of the terminal device 100 (100A) is not to end (step ST1170 "NO"), the terminal device 100 (100A) proceeds to the processing of step ST1100 and repeats the processing from step ST1100. If the terminal-side control unit (not shown) determines that the processing of the terminal device 100 (100A) is to end (step ST1170 "YES"), the terminal device 100 (100A) then ends the processing shown in FIG. 5 ("End").

[0049] Here, a detailed example of the process from the global evaluation process to the local model update process in the terminal device according to the first embodiment of the present disclosure will be described. Fig. 6 is a flowchart showing a detailed example of the process from the global evaluation process to the local model update process in the terminal device according to the first embodiment of the present disclosure. The terminal device 100 (100A) executes the process shown in Fig. 6 ("Start"), for example, from the global evaluation process to the local model update process.

[0050] The terminal device 100 (100A) then executes a global model evaluation value calculation process (step ST1210). In the global model evaluation value calculation process, the global model evaluation unit 120 of the terminal device 100 (100A) calculates an evaluation value of the global model using a data set as evaluation data. The global model evaluation unit 120 also calculates an evaluation value of the local model using a data set as evaluation data.

[0051] The terminal device 100 (100A) then executes a global model performance determination process (step ST1220 "Global model evaluation value>Local model evaluation value?"). In the global model performance determination process, the global model evaluation unit 120 of the terminal device 100 (100A) determines whether the performance of the global model is higher than the performance of the local model by determining whether the global model evaluation value is higher than the local model evaluation value.

[0052] If the global model evaluation unit 120 of the terminal device 100 (100A) determines in the global model performance determination process that the performance of the global model is higher than that of the local model ("YES" in step ST1220), the terminal device 100 (100A) then determines to adopt the global model (step ST1230). In this process, the local model update unit 130 of the terminal device 100 (100A) determines an update method for replacing the local model with the global model.

[0053] The terminal device 100 (100A) then executes a process of converting the global model into a new local model (step ST1240). In this process, the local model update unit 130 of the terminal device 100 (100A) replaces the local model with the global model and stores it in the local model storage unit 140.

[0054] If the global model evaluation unit 120 of the terminal device 100 (100A) determines in the global model performance determination process that the performance of the global model is lower than that of the local model (step ST1220 "NO"), the terminal device 100 (100A) then executes an individualization update process (step ST1250). In the individualization update process, the local model update unit 130 of the terminal device 100 (100A) performs, for example, total rejection (client-only learning) and leaves the local model as it is.

[0055] After the terminal device 100 (100A) executes the process of converting the global model into a new local model (step ST1240) or the individualization update process (step ST1250), it then ends the process shown in Fig. 6 ("End"). After the terminal device 100 (100A) ends the process shown in Fig. 6, it proceeds to the data collection process shown in Fig. 5, for example.

[0056] Next, an example of processing in a federated learning system 1 (1A) including a terminal device 100 (100A) and a server 300 according to the first embodiment of the present disclosure will be described. FIG. 7 is a sequence diagram showing an example of processing in a federated learning system 1 (1A) including a terminal device 100 (100A) and a server 300 according to the first embodiment of the present disclosure. The federated learning system 1 (1A) executes the federated learning processing shown in FIG. 7 when, for example, the server 300 acquires a global model (initial global model) that has been registered in advance as the initial value of the common model. The server 300 of the federated learning system 1 (1A) then executes a global model information distribution processing (step ST1000). In the global model information distribution processing, the global model distribution unit 330 of the server 300 outputs the global model to the terminal devices 100 (100A-1, ..., 100A-n (n≧1)).

[0057] The terminal device 100 (100A) of the federated learning system 1 (1A) then executes a global model reception process (step ST1100). In the global model reception process, the global model reception unit 110 of the terminal device 100 (100A) receives an updated global model by integrating multiple local models trained in different terminal devices (100A-1, ..., 100A-n (n ≥ 1)).

[0058] The terminal device 100 (100A) of the federated learning system 1 (1A) then executes a global model evaluation process (step ST1110). In the global model evaluation process, the global model evaluation unit 120 of the terminal device 100 (100A) evaluates the global model accepted by the global model acceptance unit 110 and the currently held local model.

[0059] The terminal device 100 (100A) of the federated learning system 1 (1A) then executes an update method determination process according to the evaluation result (step ST1120). In the update method determination process according to the evaluation result, the global model evaluation unit 120 of the terminal device 100 (100A) determines whether to adopt the global model according to the evaluation result.

[0060] The terminal device 100 (100A) of the federated learning system 1 (1A) then executes a local model update process (step ST1130). In the local model update process, the local model update unit 130 of the terminal device 100 (100A) determines a local model update method based on the evaluation result by the global model evaluation unit 120, and updates the held local model according to the determined update method. The local model holding unit 140 holds the local model updated by the local model update unit 130.

[0061] The terminal device 100 (100A) of the federated learning system 1 (1A) then executes a data collection process (step ST1140). In the data collection process, the data collection unit 150 of the terminal device 100 (100A) collects, from the information source device 200, learning data for training the local model stored in the local model storage unit 140.

[0062] The terminal device 100 (100A) of the federated learning system 1 (1A) then executes a local model learning process (step ST1150). In the local model learning process, the local model learning unit 160 of the terminal device 100 (100A) learns and constructs the local model stored in the local model storage unit 140 based on the learning data collected by the data collection unit 150.

[0063] The terminal device 100 (100A) of the federated learning system 1 (1A) then executes a local model information output process (step ST1160). In the local model information output process, the local model information output unit 170 of the terminal device 100 (100A) outputs the trained local model, which is the local model trained by the local model training unit 160, to the server 300 outside the device.

[0064] Next, the server 300 of the federated learning system 1 (1A) executes a local model information reception process (step ST1010). In the local model information reception process, the local model information reception unit 310 of the server 300 receives the trained local model trained in the terminal device 100 (100A).

[0065] The server 300 of the federated learning system 1 (1A) then executes a process of integrating the local model information into a global model (step ST1020). In this process, the global model integration unit 320 of the server 300 integrates the multiple trained local models trained in the different terminal devices (100A-1, ..., 100A-n (n ≥ 2)) to update the global model.

[0066] The federated learning system 1 (1A) then proceeds to the processing of step ST1000, where the global model distribution unit 330 outputs the global model updated by the global model integration unit 320 to the terminal devices 100 (100A-1, ..., 100A-n (n≧1)). Thereafter, the federated learning processing shown in FIG. 7 is repeatedly executed.

[0067] Next, an example of the effect of the present disclosure will be described. In federated learning technology, for example, in a video surveillance service, data accumulates daily with continuous use, new clients join, and the environment of the usage location changes (such as changing lighting), which can cause changes in distribution similarity. Furthermore, there is a risk that training data can be inferred from the transformation features used in distribution analysis, which poses a problem from the perspective of data protection. In contrast, the disclosed technology described above evaluates the global model and determines whether it should be adopted each time the global model is received from the server, allowing for dynamic switching to the optimal learning method in response to changes in the collection status of client data and other clients. Furthermore, the global model evaluation unit evaluates the global model and determines whether it should be adopted based on the model performance of the data collected by the client. The evaluation process is closed to the client, eliminating the need to share information that could be used to infer data with external parties, thereby eliminating the risk of data leakage.

[0068] This embodiment has shown an embodiment including the following configuration: [1] A terminal device including: a global model reception unit that receives an updated global model by integrating multiple trained local models, each of which is a local model trained in a different terminal device; a global model evaluation unit that evaluates the global model received by the global model reception unit and a currently held local model and determines whether to adopt the global model based on evaluation results obtained by evaluating the global model received by the global model reception unit and a currently held local model; and a local model update unit that determines a local model update method based on the evaluation results by the global model evaluation unit and updates the currently held local model in accordance with the determined update method. This makes it possible to provide a terminal device that enables an improvement in the accuracy of a learning model for a client compared to conventional federated learning.

[0069] This embodiment has described an embodiment including the following configuration: [9] A model updating method for updating a local model used in a terminal device based on a global model received in the terminal device from outside the device, wherein a global model receiving unit of the terminal device receives an updated global model by integrating multiple trained local models, each of which is a local model trained in a different terminal device; a global model evaluation unit of the terminal device evaluates the global model received by the global model receiving unit and a currently held local model, and determines whether to adopt the global model based on evaluation results of the global model receiving unit and a currently held local model; and a local model updating unit of the terminal device determines a local model update method based on the evaluation results of the global model evaluation unit, and updates the held local model in accordance with the determined update method. This makes it possible to provide a model updating method that enables improved accuracy of a learned model for a client compared to conventional federated learning.

[0070] This embodiment describes an embodiment including the following configuration: [(1)] A program causing a computer to operate as a terminal device including: a global model receiving unit that receives a global model updated by integrating multiple trained local models, each of which is a local model trained in a different terminal device; a global model evaluation unit that evaluates the global model received by the global model receiving unit and a currently held local model and determines whether to adopt the global model based on evaluation results obtained by evaluating the global model received by the global model evaluation unit and a currently held local model; and a local model update unit that determines a local model update method based on the evaluation results by the global model evaluation unit and updates the currently held local model in accordance with the determined update method. This provides an effect of providing a program that enables improved accuracy of a learning model for a client compared to conventional federated learning.

[0071] This embodiment shows an embodiment including the following configuration: [(1)] A federated learning system that performs federated learning using a local model on a terminal device side and a global model on a server side, wherein the terminal device comprises: a global model receiving unit that receives a global model updated by integrating multiple local models learned on different terminal devices; a global model evaluation unit that determines whether to adopt the global model based on evaluation results of the global model received by the global model receiving unit and a currently held local model; a local model updating unit that determines a local model update method based on the evaluation results by the global model evaluation unit and updates the held local model in accordance with the determined update method; a local model holding unit that holds the local model updated by the local model updating unit; a data collection unit that collects learning data for learning the local model held in the local model holding unit; a local model learning unit that learns and constructs the local model held in the local model holding unit based on the learning data collected by the data collection unit; and a local model information output unit that outputs a learned local model that is the local model learned by the local model learning unit to the server outside the device, and the server A federated learning system comprising: a local model information receiving unit that receives the trained local model trained on the terminal device, a global model integration unit that integrates the trained local models trained on different terminal devices to update a global model, and a global model distribution unit that outputs the global model updated by the global model integration unit to the terminal device. This makes it possible to provide a federated learning system that enables improved accuracy of the learning model for a client compared to conventional federated learning.

[0072] This embodiment further illustrates an example embodiment including the following configuration. [2] The terminal device according to [1] above, comprising: a local model holding unit that holds a local model updated by the local model update unit; a data collection unit that collects learning data for training the local model held in the local model holding unit; and a local model learning unit that builds the local model by training it based on the learning data collected by the data collection unit. This provides an advantage of providing a terminal device that can further learn and individualize a common model, such as a global model, thereby enabling improved accuracy of the learning model for a client compared to conventional federated learning. Furthermore, the present disclosure provides the same advantage as the above by applying the above configuration to a federated learning system including a terminal device, the model update method, or the program.

[0073] Embodiment 2 In embodiment 2, a configuration example of a form in which an update method to be partially adopted (individualized federated learning) is determined based on the evaluation results of the global model will be described. In embodiment 2, among the components according to embodiment 2, components that are similar to the components according to embodiment 1 already described will be given the same names and reference numerals, and duplicated descriptions will be omitted as appropriate.

[0074] Next, a configuration example of a federated learning system including a terminal device according to the second embodiment of the present disclosure will be described. FIG. 8 is a diagram illustrating an example of the configuration of a federated learning system 1 (1B) including a terminal device 100 (100B) according to the second embodiment of the present disclosure. The federated learning system 1 (1B) is a system that performs federated learning using a local model on the terminal device 100 (100B) side and a global model on the server 300 side. The federated learning system 1 (1B) collects the local model on the terminal device 100 (100B) side, updates the global model using the local model, and distributes the global model to the terminal device 100 (100B). The federated learning system 1 (1B) shown in FIG. 8 is configured to include the terminal device 100 (100B), an information source device 200, and a server 300.

[0075] The terminal device 100 (100B) is configured to have a different local model update method than the terminal device 100 (100A) already described. In the federated learning system 1 (1B), the terminal device 100 (100B), like the terminal device 100 (100A) already described, may correspond to each of multiple terminal devices 100 (100B-1, 100B-2, 100B-3, ..., 100B-n). The terminal device 100 (100B) shown in FIG. 8 is configured to include a global model acceptance unit 110, a global model evaluation unit 120, a local model update unit 130 (130B), a local model holding unit 140, a data collection unit 150, a local model learning unit 160, and a local model information output unit 170. The global model acceptance unit 110 has the same functions and is configured similarly to the global model acceptance unit 110 already described. The global model evaluation unit 120 has the same functions and is configured in the same manner as the global model evaluation unit 120 already described.

[0076] When the evaluation result by the global model evaluation unit 120 indicates that the performance of the global model accepted by the global model acceptance unit 110 is lower than that of the currently held local model, the local model update unit 130 (130B) combines a portion of the held local model with a portion of the global model to determine an update method for updating the held local model. Alternatively, when the evaluation result by the global model evaluation unit 120 indicates that the performance of the global model accepted by the global model acceptance unit 110 is lower than that of the currently held local model, the local model update unit 130 (130B) determines the update method by gradually changing the combination of a portion of the held local model with a portion of the global model. The model update method determined by the local model update unit 130 (130B) will be described. FIG. 9 is a diagram for explaining a model update method by a terminal device according to a second embodiment of the present disclosure. FIG. 9 illustrates an image of a model (local model) 2000. In the terminal device 100 (100B), when a global model is received and the performance of the global model is found to be lower than the performance of the model (local model) 2000, the model update method is determined to be an update method in which the parameters of the local model are used as the parameters of the node connecting the final layer 2400 in the hidden layer of the model (local model) 2000 and the output layer 2500, among the layers from the input layer 2100, which is the input data, to the output layer 2500. In this way, an update method is determined to combine a portion of the held local model with a portion of the global model and update the held local model. Furthermore, an update method in which the parameters of the node connected before the final layer 2400 are used as the parameters of the local model in a stepwise manner may also be determined.

[0077] The local model holding unit 140 has the same functions as the already-described local model holding unit 140 and is configured similarly. The data collection unit 150 has the same functions as the already-described data collection unit 150 and is configured similarly. The local model learning unit 160 has the same functions as the already-described local model learning unit 160 and is configured similarly. The local model information output unit 170 has the same functions as the already-described local model information output unit 170 and is configured similarly.

[0078] The information source device 200 outputs data to be used in the terminal device 100 (100B), similar to the information source device 200 already described. In the federated learning system 1 (1B), an information source device 200 is provided for each client, and the information source device 200 can correspond to each of the multiple information source devices 200-1, 200-2, 200-3, ..., 200-n.

[0079] Similar to the server 300 already described, the server 300 receives local models from the client terminal devices 100 (100B), integrates multiple local models to update the global model, and distributes the updated global model to each of the terminal devices 100 (100B). Similar to the server 300 already described, the server 300 includes a local model information receiving unit 310, a global model integration unit 320, and a global model distribution unit 330.

[0080] Next, a processing example of a terminal device according to the second embodiment of the present disclosure will be described. FIG. 10 is a flowchart illustrating an example of a personalization update process in a terminal device according to the second embodiment of the present disclosure. The terminal device 100 (100B) illustrated in FIG. 8 executes the process illustrated in FIG. 10 in the personalization update process ("Start"). The terminal device 100 (100B) then executes a personalization parameter determination process (step ST2251). In the personalization parameter determination process, the local model update unit 130 (130B) of the terminal device 100 (100C) determines parameters that are considered to have good performance among the parameters constituting the local model as personalization parameters. The terminal device 100 (100B) then executes an update process for shared parameters other than the personalization parameters based on the global model (step ST2252). In this update process, the local model update unit 130 (130B) of the terminal device 100 (100B) replaces parameters in the local model other than the personalization parameters with parameters of the global model. Here, in the update process, the local model update unit 130 (130B) may first set the parameters of the final layer 2400 as individualized parameters, replace all parameters other than the individualized parameters with parameters of the global model, evaluate the replaced model, and determine the update method by gradually changing the parameters of each layer in turn to individualized parameters based on the evaluation results. After executing the process of step ST1252, the local model update unit 130 (130B) of the terminal device 100 (100B) then terminates the process shown in FIG. 10 ("End"). In this way, when the evaluation result by the global model evaluation unit 120 indicates that the performance of the global model accepted by the global model acceptance unit 110 is lower than the performance of the currently held local model, the local model update unit 130 (130B) combines a portion of the held local model with a portion of the global model to determine an update method for updating the held local model.Alternatively, if the evaluation result by the global model evaluation unit 120 indicates that the performance of the global model accepted by the global model acceptance unit 110 is lower than the performance of the currently held local model, the local model update unit 130 (130B) determines the update method by gradually changing the combination of a portion of the held local model and a portion of the global model.

[0081] Next, an example of processing in a federated learning system including a terminal device and a server according to embodiment 2 of the present disclosure will be described. Fig. 11 is a flowchart showing an example of processing in a federated learning system including a terminal device and a server according to embodiment 2 of the present disclosure. The federated learning system 1 (1B) shown in Fig. 8 executes the federated learning processing shown in Fig. 11 when, for example, the server 300 acquires a global model (initial global model) that has been registered in advance as the initial value of the common model.

[0082] Next, the server 300 of the federated learning system 1 (1B) executes a global model distribution process (step ST2100). In this process, the global model distribution unit 330 of the server 300 distributes the global model to the clients.

[0083] The terminal device 100 (100B) of the federated learning system 1 (1B) then executes a process (step ST2110). In this process, the global model receiving unit 110 of the terminal device 100 (100B) receives the global model.

[0084] The terminal device 100 (100B) of the federated learning system 1 (1B) then executes a process (step ST2120). In this process, the global model evaluation unit 120 of the terminal device 100 (100B) evaluates the performance of the global model on the client data.

[0085] The terminal device 100 (100B) of the federated learning system 1 (1B) then executes a global model performance evaluation process, in which the global model evaluation unit 120 of the terminal device 100 (100B) determines whether the performance of the global model is higher than the performance of the local model after the previous update for all classes (step ST2130).

[0086] When the terminal device 100 (100B) of the federated learning system 1 (1B) determines that the performance of the global model is higher than that of the local model after the previous update for all classes ("YES" in step ST2130), it then executes a local model update process (step ST2140). In this process, the local model update unit 130 (130B) of the terminal device 100 (100B) copies all parameters of the global model to the local model.

[0087] When the terminal device 100 (100B) of the federated learning system 1 (1B) determines that the performance of the global model is not higher than that of the local model after the previous update for at least one or more classes (step ST2130 "NO"), the terminal device 100 (100B) then executes a personalization parameter determination process. In this process, the local model update unit 130 (130B) of the terminal device 100 (100B) expands the personalization parameters (step ST2150).

[0088] The terminal device 100 (100B) of the federated learning system 1 (1B) then executes a local model update process (step ST2160). In this process, the local model update unit 130 (130B) of the terminal device 100 (100B) copies only the shared parameters from the global model to the local model.

[0089] The terminal device 100 (100B) of the federated learning system 1 (1B) then executes a local model learning process. In this process, the local model learning unit 160 of the terminal device 100 (100B) learns and constructs a local model using the client data (creating a local model) (step ST2170).

[0090] The terminal device 100 (100B) of the federated learning system 1 (1B) then executes a local model information output process. In this process, the local model information output unit 170 of the terminal device 100 (100B) transmits the local model to the server (step ST2180).

[0091] The server 300 of the federated learning system 1 (1B) then executes a process of integrating the local models collected from multiple clients to create a global model (step ST2190). In this process, the local model information receiving unit 310 of the server 300 collects local model information from multiple clients. Next, the global model integration unit 320 of the server 300 integrates the collected local models to update the global model. The federated learning system 1 (1B) then executes an end determination process (step ST2200 "End?"). If the federated learning system 1 (1B) determines not to end the process in the end determination process (step ST2200 "End?" "NO"), it then proceeds to the process of step ST2100 and repeats the process from step ST2100. If the federated learning system 1 (1B) determines to end the process in the end determination process (step ST2200 "End?" "YES"), it then ends the process shown in FIG. 11 ("End").

[0092] Next, the effects of the second embodiment of the present disclosure will be described. In the present disclosure, when a global model is received from the server, in the case of normal federated learning, the global model is immediately adopted as the local model. In the present invention, the global model evaluation means allows the user to select the level of adoption of the global model from options such as full adoption (normal federated learning), partial adoption (individualized federated learning), or full rejection (client-only learning). In partial adoption (individualized federated learning), if the performance of the global model is lower than that of the local model, only some of the model's parameters are adopted, and the remaining parameters are uniqueized. If unique parameters already exist, the number of unique parameters can be increased. In federated learning, the effectiveness of learning varies greatly depending on the distribution and amount of learning data for each client. The disclosed technology allows a client to participate in federated learning if it is effective, and gradually withdraw from federated learning if it is not. This prevents a decline in local performance and optimizes the benefits of federated learning. Because the above determination is made each time a global model is received from the server, the learning method can be dynamically switched to the optimal method depending on changes in the collection status of client data for the client and other clients. Global model evaluation is based on the model performance of the client's own collected data. The evaluation process is closed to the client, eliminating the need to share information that could be used to infer data with external parties, and eliminating the risk of data leaks.

[0093] This embodiment further illustrates an exemplary embodiment including the following configuration. [3] The terminal device according to [1] or [2], characterized in that, when the evaluation result by the global model evaluation unit indicates that the performance of the global model accepted by the global model acceptance unit is lower than the performance of the currently held local model, the local model update unit determines an update method for updating the held local model by combining a portion of the held local model with a portion of the global model. This further provides an effect of providing a terminal device that, by combining a local model and a global model, enables the accuracy of a learning model for a client to be improved compared to conventional federated learning. Furthermore, the present disclosure provides the same effect as the above by applying the above configuration to a federated learning system including a terminal device, the model update method, or the program.

[0094] This embodiment further illustrates an exemplary embodiment including the following configuration. [4] The terminal device according to [1] or [2], characterized in that, when the evaluation result by the global model evaluation unit indicates that the performance of the global model accepted by the global model acceptance unit is lower than the performance of the currently held local model, the local model update unit determines the update method by gradually changing the combination of a portion of the held local model and a portion of the global model. This further provides an effect of providing a terminal device that can prevent, for example, a sudden change in the output of the local model after updating, thereby enabling further improvement in the accuracy of the learning model for the client compared to conventional federated learning. Furthermore, the present disclosure provides the same effect as the above by applying the above configuration to a federated learning system including a terminal device, the model update method, or the program.

[0095] Embodiment 3. In the above-described embodiment, a configuration example has been described in which a local model used in a client-side terminal device can be personalized for each client. If the personalized local model is used as is in federated learning, the accuracy of the global model as a common model may deteriorate. In embodiment 3, a configuration example will be described in which the personalized situation can be taken into consideration. In embodiment 3, among the components according to embodiment 3, components that are similar to the components according to embodiment 1 or embodiment 2 already described will be given the same names and similar reference numerals, and duplicate explanations will be omitted as appropriate.

[0096] Next, a configuration example of a federated learning system including a terminal device according to the third embodiment of the present disclosure will be described. FIG. 12 is a diagram illustrating an example of the configuration of a federated learning system including a terminal device according to the third embodiment of the present disclosure. The federated learning system 1 (1C) performs federated learning using a local model on the terminal device 100 (100C) side and a global model on the server 300 side. The federated learning system 1 (1C) collects local models on the terminal device 100 (100C) side, updates the global model using the local models, and distributes the global model to the terminal device 100 (100C). The federated learning system 1 (1C) is further configured to update the global model taking into account the individualization status of the local model. The federated learning system 1 (1C) shown in FIG. 12 includes a terminal device 100 (100C), an information source device 200, and a server 300 (300C).

[0097] The terminal device 100 (100C) further has a function of calculating the individualization status of the local model and outputting it to an external device (server 300). The terminal device 100 (100C) shown in FIG. 12 includes a global model receiving unit 110, a global model evaluating unit 120, a local model updating unit 130 (130C), a local model holding unit 140, a data collecting unit 150, a local model learning unit 160, a local model information output unit 170 (170C), and an individualization status calculating unit 180.

[0098] The global model receiving unit 110 receives a global model that has been updated by integrating a plurality of local models that have been trained on different terminal devices.

[0099] The global model evaluation unit 120 evaluates the global model received by the global model reception unit 110 and the currently held local model, and determines whether to adopt the global model based on the evaluation result. The evaluation result is output in the form of, for example, an evaluation value.

[0100] The local model update unit 130 (130C) determines a method for updating the local model based on the evaluation result by the global model evaluation unit 120, and updates the held local model in accordance with the determined update method.

[0101] The personalization status calculation unit 180 calculates a personalization status indicating the degree to which the local model updated based on the update method determined by the local model update unit 130 (130C) is personalized relative to the global model. The parameter personalization status of the global model is, for example, the ratio of personalized parameters to all parameters of the model. In this case, it can also be expressed as the ratio of parameters that do not adopt the global model.

[0102] The local model holding unit 140 holds the local model updated by the local model update unit 130 (130C).

[0103] The data collection unit 150 collects learning data for training the local model held in the local model holding unit.

[0104] The local model learning unit 160 constructs the local model held in the local model holding unit by learning it based on the learning data collected by the data collecting unit.

[0105] Like the local model information output unit 170 already described, the local model information output unit 170 (170C) has a function of outputting, to the outside of the device, a trained local model, which is a local model trained by the local model learning unit 160. Furthermore, the local model information output unit 170 (170C) outputs, to the outside of the device, the individualization status together with the trained local model, which is a local model trained by the local model learning unit 160. Specifically, the local model information output unit 170 (170C) outputs the trained local model, which is a local model trained by the local model learning unit 160, and the individualization status to the server 300 outside the device.

[0106] The information source device 200 outputs data to be used in the terminal device 100 (100C), similar to the information source device 200 already described. In the federated learning system 1 (1C), an information source device 200 is provided for each client, and the information source device 200 can correspond to each of the multiple information source devices 200-1, 200-2, 200-3, ..., 200-n.

[0107] Similar to the server 300 already described, the server 300 (300C) has the functions of receiving local models from the client terminal devices 100 (100C), integrating multiple local models to update the global model, and distributing the updated global model to each of the terminal devices 100 (100B). Furthermore, when integrating multiple local models to update the global model, the server 300 (300C) takes into account the individualization status during the update. The server 300 (300C) shown in FIG. 12 includes a local model information receiving unit 310 (310C), a global model integration unit 320 (320C), and a global model distribution unit 330 (330C).

[0108] The local model information receiving unit 310 (310C) receives a local model learned in the terminal device 100 (100C) and an individualization status indicating the degree to which the local model is individualized with respect to the global model.

[0109] The global model integration unit 320 (320C) updates the global model by integrating multiple trained local models, which are the local models trained in different terminal devices 100 (100C), according to the individualization status of each trained local model. When creating a global model as an integrated model, the global model integration unit 320 (320C) can take the average of all updated models, but it can also take a weighted average, for example, using the parameter individualization status as a weight. That is, the global model integration unit 320 (320C) determines weights for parameter integration based on the individualization rate (uniqueness rate) of parameters in the local model on the client side, and performs weighting and integration.

[0110] The global model distribution unit 330 (330C) outputs the global model updated by the global model integration unit 320 (320C) to the terminal device 100 (100C).

[0111] Next, a processing example of a terminal device according to embodiment 3 of the present disclosure will be described. Fig. 13 is a flowchart showing an example of processing in a federated learning system including a terminal device and a server according to embodiment 3 of the present disclosure. Fig. 14 is a sequence diagram showing an example of processing in a federated learning system including a terminal device and a server according to embodiment 3 of the present disclosure. The federated learning system 1 (1C) shown in Fig. 12 executes the federated learning processing shown in Figs. 13 and 14 when, for example, the server 300 acquires a global model (initial global model) registered in advance as the initial value of the common model.

[0112] Next, the server 300 of the federated learning system 1 (1C) executes a global model distribution process (step ST3000). In the global model distribution process, the global model distribution unit 330 (330C) of the server 300 outputs the global model updated by the global model integration unit 320 (320C) to the terminal device 100 (100C).

[0113] The terminal device 100 (100C) of the federated learning system 1 (1C) starts processing when it receives a global model from the server 300. The terminal device 100 (100C) executes a global model reception process (step ST3100). In the global model reception process, the global model reception unit 110 of the terminal device 100 (100C) executes the global model reception process.

[0114] The terminal device 100 (100C) of the federated learning system 1 (1C) then executes a global model evaluation process (step ST3110). The global model evaluation process is executed by the global model evaluation unit 120 of the terminal device 100 (100C).

[0115] The terminal device 100 (100C) of the federated learning system 1 (1C) then executes an update method determination process according to the evaluation result (step ST3120). The update method determination process according to the evaluation result is executed by the local model update unit 130 (130C) of the terminal device 100 (100C).

[0116] The terminal device 100 (100C) of the federated learning system 1 (1C) then executes a local model update process (step ST3130). In the local model update process, the local model update unit 130 (130C) of the terminal device 100 (100C) executes the local model update process in the same manner as the local model update process already described.

[0117] The terminal device 100 (100C) of the federated learning system 1 (1C) then executes an individualization status calculation process (step ST3140). In the individualization status calculation process, the individualization status calculation unit 180 of the terminal device 100 (100C) calculates an individualization status indicating the degree to which the local model updated based on the update method determined by the local model update unit 130 (130C) is individualized with respect to the global model.

[0118] The terminal device 100 (100C) of the federated learning system 1 (1C) then executes a data collection process (step ST3150). In the data collection process, the data collection unit 150 of the terminal device 100 (100C) executes the process in the same manner as the local model update process already described.

[0119] The terminal device 100 (100C) of the federated learning system 1 (1C) then executes a local model learning process (step ST3160). In the local model learning process, the local model learning unit 160 of the terminal device 100 (100C) executes the local model update process in the same manner as the local model update process already described.

[0120] The terminal device 100 (100C) of the federated learning system 1 (1C) then executes a local model information and individualization status output process (step ST3170). In this process, the local model information output unit 170 (170C) of the terminal device 100 (100C) outputs the individualization status to the outside of the device together with the trained local model, which is the local model trained by the local model training unit 160. Specifically, the local model information output unit 170 (170C) outputs the trained local model, which is the local model trained by the local model training unit 160, and the individualization status to the server 300 outside the device.

[0121] Server 300 of federated learning system 1 (1C) then executes a process for receiving local model information and an individualization status (step ST3200). In the process, local model information receiving unit 310 (310C) of server 300 receives the local model learned in terminal device 100 (100C) and an individualization status indicating the degree to which the local model is individualized with respect to the global model.

[0122] The server 300 of the federated learning system 1 (1C) then executes a global model integration process (step ST3210). In the global model integration process, the global model integration unit 320 (320C) of the server 300 integrates multiple trained local models, which are the local models trained in the different terminal devices 100 (100C), according to the individualization status of each trained local model, to update the global model. The global model integration unit 320 (320C) executes a process of weighting local model information according to the individualization status and integrating the weighted local model information into the global model.

[0123] The associative learning system 1 (1C) then executes an end determination process (step ST3300 "End?").

[0124] If the associative learning system 1 (1C) determines not to terminate the processing in the termination determination process (step ST3300 "End?" "NO"), it then proceeds to the processing of step ST3000 and repeats the processing from step ST3000. Thereafter, the associative learning system 1 (1C) proceeds to the processing of, for example, step ST3000, and repeatedly executes the associative learning processing shown in Figures 13 and 14.

[0125] When it is determined in the termination determination process that the process is to be terminated (step ST3300 "End?" "YES"), the associative learning system 1 (1C) then terminates the process shown in FIG. 13 ("End").

[0126] Next, the effects of the third embodiment of the present disclosure will be described. In the present disclosure, a configuration has been shown in which, when a client transmits an updated model to a server, the client also notifies the server of the parameter personalization status (uniqueness status). Furthermore, when creating a global model, a configuration has been shown in which weights for creating an integrated model are determined in accordance with the client's parameter personalization status (uniqueness status). This makes it possible to generate a global model by integrating multiple local models according to the personalization status of the local models.

[0127] This embodiment further illustrates an example embodiment including the following configuration. [5] The terminal device according to [2] above, characterized in that it includes: an individualization status calculation unit that calculates an individualization status indicating the degree to which a local model updated based on an update method determined by the local model update unit is individualized relative to the global model; and a local model information output unit that outputs the individualization status to an external device together with a trained local model, which is the local model trained by the local model training unit. This further provides an advantage of providing a terminal device that can take the degree of individualization of a local model into consideration when updating a global model, thereby enabling the accuracy of a learned model for a client to be further improved compared to conventional federated learning. Furthermore, the present disclosure provides the same advantage as the above by applying the above configuration to a federated learning system including a terminal device, the model update method, or the program.

[0128] This embodiment shows an embodiment including the following configuration: [7] A federated learning system that performs federated learning using a local model on a terminal device side and a global model on a server side, wherein the terminal device comprises: a global model receiving unit that receives a global model updated by integrating a plurality of local models learned on different terminal devices; a global model evaluation unit that determines whether to adopt the global model based on evaluation results of the global model received by the global model receiving unit and a currently held local model; a local model updating unit that determines a local model update method based on the evaluation results by the global model evaluation unit and updates the held local model in accordance with the determined update method; a local model holding unit that holds the local model updated by the local model updating unit; a data collection unit that collects learning data for learning the local model held in the local model holding unit; a local model learning unit that learns and constructs the local model held in the local model holding unit based on the learning data collected by the data collection unit; and an individualization status calculation unit that calculates an individualization status indicating the degree to which the local model updated based on the update method determined by the local model updating unit is individualized with respect to the global model. a local model information output unit that outputs a trained local model, which is a local model trained by the local model learning unit, and the individualization status to the server outside the device, wherein the server comprises: a local model information receiving unit that receives the trained local model trained on the terminal device and the individualization status; a global model integration unit that updates a global model by integrating multiple trained local models trained on different terminal devices according to the individualization status of each trained local model; and a global model distribution unit that outputs the global model updated by the global model integration unit to the terminal device.As a result, the present disclosure has the effect of providing a federated learning system that makes it possible to take into account the degree of individualization of local models when updating the global model, thereby enabling the accuracy of the federated learning model for the client to be further improved compared to conventional federated learning.

[0129] This embodiment has shown an embodiment including the following configuration: [8] A server comprising: a local model information receiving unit that receives a local model trained in a terminal device and an individualization status indicating the degree to which the local model is individualized with respect to a global model; a global model integrating unit that updates a global model by integrating multiple trained local models, each of which is the local model trained in a different terminal device, in accordance with the individualization status of each trained local model; and a global model distribution unit that outputs the global model updated by the global model integrating unit to the terminal device. This makes it possible to take the degree of individualization of a local model into consideration when updating a global model, thereby providing an effect of providing a server that enables the accuracy of a federated learning model for a client to be improved compared to conventional federated learning.

[0130] Fourth Embodiment When there are multiple types (classes) of model output data, the accuracy may differ for each class. In the fourth embodiment, a configuration example will be described in which an update method is determined using the evaluation results for each class. In the fourth embodiment, among the components according to the fourth embodiment, components similar to those according to the first, second, or third embodiment already described will be assigned the same names and the same reference numerals, and redundant description will be omitted as appropriate.

[0131] Next, a configuration example of a federated learning system including a terminal device according to embodiment 4 of the present disclosure will be described. FIG. 15 is a diagram illustrating an example of the configuration of a federated learning system including a terminal device according to embodiment 4 of the present disclosure. The federated learning system 1 (1D) is a system that performs federated learning using a local model on the terminal device 100 (100D) side and a global model on the server 300 side. The federated learning system 1 (1D) collects local models on the terminal device 100 (100D) side, updates the global model using the local models, and distributes the global model to the terminal device 100 (100D). The federated learning system 1 (1D) shown in FIG. 15 is configured to include a terminal device 100 (100D), an information source device 200, and a server 300 (300D).

[0132] The terminal device 100 (100D) is configured to have a different local model update method than the terminal device 100 (100C) already described. In the federated learning system 1 (1D), the terminal device 100 (100D), like the terminal device 100 (100C) already described, may correspond to each of the multiple terminal devices 100 (100D-1, 100D-2, 100D-3, ..., 100D-n). The terminal device 100 (100D) shown in FIG. 15 is configured to include a global model acceptance unit 110, a global model evaluation unit 120 (120D), a local model update unit 130 (130D), a local model holding unit 140, a data collection unit 150, a local model learning unit 160, a local model information output unit 170 (170D), and an individualization status calculation unit 180.

[0133] The global model receiving unit 110 has the same functions and is configured in the same manner as the global model receiving unit 110 already described.

[0134] The global model evaluation unit 120 (120D) has the same functions and is configured as any of the already-described global model evaluation units 120. Furthermore, the global model evaluation unit 120 (120D) performs evaluation for each class that can be output by the global model accepted by the global model acceptance unit 110 and the currently held local model.

[0135] The local model updating unit 130 (130D) has the same functions and is configured as any of the local model updating units 130 already described. Furthermore, the local model updating unit 130 (130D) determines an update method for each class using the evaluation results for each class by the global model evaluation unit 120 (120D). The model update method determined by the local model updating unit 130 (130D) will be described. Figure 16 is a diagram for explaining a model update method by a terminal device according to embodiment 4 of the present disclosure. Figure 16 shows an image of a model (local model) 4000. In the terminal device 100 (100D), when it is determined that the performance of the global model is lower than the performance of the model (local model) 4000 in class C 4600 of the output layer 4500, the model update method is determined to be an update method in which the parameters of the node 4600 connected to class C in the output layer 4500, among the parameters of the node connecting the final layer 4400 in the hidden layer of the layers from the input layer 4100 that is the input data to the output layer 4500, are set as parameters of the local model. In this way, an update method is determined to combine a portion of the held local model with a portion of the global model to update the held local model. Furthermore, an update method is determined to adopt the parameters of the node connected to a class in the final layer 4500 for which the local model has better performance as parameters of the local model.

[0136] The local model holding unit 140 has the same functions and is configured in the same manner as the local model holding unit 140 already described.

[0137] The data collection unit 150 has the same functions and is configured in the same manner as the data collection unit 150 already described.

[0138] The local model learning unit 160 has the same functions and is configured in the same manner as the local model learning unit 160 already described.

[0139] The local model information output unit 170 (170D) has the same functions and configuration as the already-described local model information output unit 170 (170C). The individualization status calculation unit 180 has the same functions and configuration as the already-described individualization status calculation unit 180.

[0140] The information source device 200 outputs data to be used in the terminal device 100 (100D), similar to the information source device 200 already described. In the federated learning system 1 (1D), an information source device 200 is provided for each client, and the information source device 200 can correspond to each of the multiple information source devices 200-1, 200-2, 200-3, ..., 200-n.

[0141] Like the server 300 already described, the server 300 (300D) has the functions of accepting local models from the client terminal devices 100 (100D), integrating multiple local models to update a global model, and distributing the updated global model to each of the terminal devices 100 (100D). Furthermore, when integrating multiple local models to update the global model, the server 300 (300D) takes into account the individualization status. The server 300 (300D) shown in FIG. 15 includes a local model information accepting unit 310 (310D), a global model integrating unit 320 (320D), and a global model distributing unit 330 (330D). The local model information accepting unit 310 (310D) has the same functions as the local model information accepting unit 310 (310C) already described. The global model integrating unit 320 (320D) has the same functions as the global model integrating unit 320 (320C) already described. The global model distribution unit 330 (330D) has the same functions as the global model distribution unit 330 (330C) already described.

[0142] Next, a processing example of the terminal device according to the fourth embodiment of the present disclosure will be described. Fig. 17 is a flowchart showing a detailed example of the process from the global evaluation process to the local model update process in the terminal device according to the fourth embodiment of the present disclosure. The terminal device 100 (100D) shown in Fig. 15 executes the process shown in Fig. 17 ("START"), for example, from the global evaluation process to the local model update process.

[0143] The terminal device 100 (100D) then executes an evaluation value calculation process. In the evaluation value calculation process, the global model evaluation unit 120 (120D) of the terminal device 100 (100D) executes the evaluation value calculation process for each class of the global model (step ST4210).

[0144] The terminal device 100 (100D) then executes a global model performance determination process. In the global model performance determination process, the global model evaluation unit 120 (120D) of the terminal device 100 (100D) executes the global model performance determination process for each class (step ST4220 "Global model evaluation value>local model evaluation value?").

[0145] The terminal device 100 (100D) then determines to adopt the global model (step ST4230). In this process, the local model update unit 130 (130D) of the terminal device 100 (100D) specifically adopts the global model for a class whose global model evaluation value is higher than the local model evaluation value.

[0146] The terminal device 100 (100D) then executes a process of converting the global model into a new local model (step ST4240). In this process, the local model update unit 130 (130D) of the terminal device 100 (100D) specifically updates the global model parameters for classes whose global model evaluation values ​​are higher than their local model evaluation values, using the global model parameters.

[0147] The terminal device 100 (100D) then executes an individualization update process (step ST4250). In the individualization update process, the local model update unit 130 (130D) of the terminal device 100 (100D) specifically adopts a local model for a class whose global model evaluation value is equal to or less than the local model evaluation value.

[0148] When the local model update unit 130 (130D) executes the process of updating the local model in the process of step ST4240 or step ST4250, the terminal device 100 (100D) then ends the process shown in Fig. 17 ("End"). When the terminal device 100 (100D) ends the process shown in Fig. 17, it proceeds to, for example, the individualization situation calculation process shown in Fig. 13.

[0149] Next, an example of the personalization update process in the terminal device according to embodiment 4 will be described. Fig. 18 is a flowchart showing an example of the personalization update process in the terminal device according to embodiment 4 of the present disclosure. The terminal device 100 (100D) shown in Fig. 15 executes the process shown in Fig. 18 in the personalization update process ("start").

[0150] The terminal device 100 (100D) then executes an individualization parameter determination process. In the individualization parameter determination process, the local model update unit 130 (130D) of the terminal device 100 (100D) executes the individualization parameter determination process in accordance with the evaluation result for each class (step ST4251).

[0151] The terminal device 100 (100D) then executes a model update process (step ST4252). In this update process, the local model update unit 130 (130D) of the terminal device 100 (100D) updates the shared parameters other than the individualization parameters based on the global model. Here, in this update process, the local model update unit 130 (130D) may first set the parameters corresponding to node 4600 in the final layer 4400 connected to class C in the output layer 4500 as individualization parameters, replace all parameters other than the individualization parameters with parameters of the global model, evaluate the replaced model, and determine an update method by gradually changing the parameters of each layer in turn based on the evaluation results.

[0152] After the local model update unit 130 (130D) executes the process of step ST4252, the terminal device 100 (100D) then ends the process shown in FIG.

[0153] Next, the effects of the fourth embodiment of the present disclosure will be described. In the present disclosure, when individualizing (uniqueizing) parameters of the final layer in a model that performs class classification (image classification, object detection, etc.), it is possible to individualize (unique) only nodes connected to classes in which the performance of the global model is lower than that of the local model, rather than individualizing (unique) parameters of the entire layer at once. In other words, when individualizing (uniqueizing) parameters, it is possible to determine the parameters to be individualized (unique) for each class, rather than for each layer.

[0154] This embodiment further illustrates an example embodiment including the following configuration. [6] The terminal device according to any one of [1] to [5] above, characterized in that the global model evaluation unit performs evaluation for each class that can be output using the global model accepted by the global model acceptance unit and the currently held local model, and the local model update unit determines an update method for each class using the evaluation result for each class by the global model evaluation unit. This provides an advantage that the present disclosure can provide a terminal device that can further improve the accuracy of the learning model for a client compared to conventional federated learning, since it can determine an update method for each class based on the evaluation result for each class. Furthermore, the present disclosure provides an advantage similar to the above advantage by applying the above configuration to a federated learning system including a terminal device, the model update method, or the program.

[0155] Here, a hardware configuration for realizing the functions of the present disclosure will be described. Fig. 19 is a diagram illustrating a first example of a hardware configuration for realizing the functions of the configuration of the present disclosure. Fig. 20 is a diagram illustrating a second example of a hardware configuration for realizing the functions of the configuration of the present disclosure. The terminal devices 100 (100A, 100B, 100C, 100D), 100-1, 100-2, 100-3, ..., 100-n, or the servers 300 (300C, 300D) of the present disclosure are each realized by hardware such as that shown in Fig. 19 or Fig. 20.

[0156] As shown in FIG. 19 , each of the terminal devices 100 (100A, 100B, 100C, 100D), 100-1, 100-2, 100-3, ..., 100-n is configured with, for example, a processor 10001, a memory 10002, an input / output interface 10003, and a communication circuit 10004. The processor 10001 and the memory 10002 are, for example, mounted on a computer. The memory 10002 stores programs that cause the computer to function as a global model receiving unit 110, a global model evaluation unit 120 (120D), a local model updating unit 130 (130B, 130C, 130D), part of the local model holding unit 140, a data collecting unit 150, a local model learning unit 160, a local model information output unit 170 (170C, 170D), an individualization status calculation unit 180, and a terminal control unit (not shown). By having the processor 10001 read and execute the programs stored in the memory 10002, the functions of the global model receiving unit 110, the global model evaluation unit 120 (120D), the local model updating unit 130 (130B, 130C, 130D), part of the local model holding unit 140, the data collecting unit 150, the local model learning unit 160, the local model information output unit 170 (170C, 170D), the individualization status calculation unit 180, and a terminal-side control unit (not shown) are realized. Furthermore, the memory 10002 or another memory (not shown) realizes part of the local model holding unit 140 and a terminal-side storage unit (not shown). Furthermore, the communication circuit 10004 realizes a terminal-side communication unit (not shown).

[0157] 19, each of the servers 300 (300C, 300D) is configured with, for example, a processor 10001, a memory 10002, an input / output interface 10003, and a communication circuit 10004. The processor 10001 and the memory 10002 are mounted on, for example, a computer. The memory 10002 stores programs that cause the computer to function as a local model information receiving unit 310 (310C, 310D), a global model integration unit 320 (320C, 320D), a global model distribution unit 330 (330C, 330D), and a server-side control unit (not shown). The processor 10001 reads and executes the programs stored in the memory 10002 to realize the functions of the local model information receiving unit 310 (310C, 310D), the global model integration unit 320 (320C, 320D), the global model distribution unit 330 (330C, 330D), and a server-side control unit (not shown). The memory 10002 or another memory (not shown) realizes a server-side storage unit (not shown). The communication circuit 10004 realizes a server-side communication unit (not shown).

[0158] The processor 10001 is, for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), a microprocessor, a microcontroller, or a DSP (Digital Signal Processor). The memory 10002 may be a non-volatile or volatile semiconductor memory such as RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable Read Only Memory), or flash memory, or a magnetic disk such as a hard disk or flexible disk, or an optical disk such as a CD (Compact Disc) or DVD (Digital Versatile Disc), or a magneto-optical disk. The processor 10001 and the memory 10002 or the communication circuit 10004 are connected in a state capable of transmitting data to each other. The processor 10001, memory 10002, and communication circuit 10004 are connected via an input / output interface 10003 so as to be capable of transmitting data to and from other hardware.

[0159] Alternatively, the functions of the global model receiving unit 110, the global model evaluation unit 120 (120D), the local model updating unit 130 (130B, 130C, 130D), part of the local model holding unit 140, the data collecting unit 150, the local model learning unit 160, the local model information output unit 170 (170C, 170D), the individualization status calculation unit 180, and the terminal-side control unit (not shown) in the terminal devices 100 (100A, 100B, 100C, 100D), 100-1, 100-2, 100-3, ..., 100-n may be realized by a dedicated processing circuit 20001, as shown in FIG. 20 .

[0160] Similarly, the functions of the local model information receiving unit 310 (310C, 310D), global model integration unit 320 (320C, 320D), global model distribution unit 330 (330C, 330D), and a server-side control unit (not shown) in the server 300 (300C, 300D) may be realized by a dedicated processing circuit 20001, as shown in FIG. 20 .

[0161] The processing circuit 20001 may be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), an FPGA (Field-Programmable Gate Array), a SoC (System-on-a-Chip), or a system LSI (Large-Scale Integration), etc. In addition, the memory 20002 or another memory (not shown) implements part of the local model holding unit 140 and a storage unit (not shown). The memory 20002 may be a non-volatile or volatile semiconductor memory such as RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable Read Only Memory), or flash memory, or a magnetic disk such as a hard disk or flexible disk, or an optical disk such as a CD (Compact Disc) or DVD (Digital Versatile Disc), or a magneto-optical disk. The communication circuit 20004 implements a communication unit (not shown). The processing circuit 20001 and the memory 20002 or the communication circuit 20004 are connected in a state where they can transmit data to each other. In addition, the processing circuit 20001, the memory 20002, and the communication circuit 20004 are connected in a state where they can transmit data to other hardware via the input / output interface 20003.In addition, in the terminal devices 100 (100A, 100B, 100C, 100D), 100-1, 100-2, 100-3, ..., 100-n, the functions of the global model receiving unit 110, the global model evaluation unit 120 (120D), the local model updating unit 130 (130B, 130C, 130D), part of the local model holding unit 140, the data collecting unit 150, the local model learning unit 160, the local model information output unit 170 (170C, 170D), the individualization status calculation unit 180, and the terminal-side control unit (not shown) may be realized by separate processing circuits, or may be realized collectively by a processing circuit. Similarly, in the server 300 (300C, 300D), the functions of the local model information receiving unit 310 (310C, 310D), the global model integration unit 320 (320C, 320D), the global model distribution unit 330 (330C, 330D), and the server-side control unit (not shown) may be realized by separate processing circuits, or may be realized collectively by a processing circuit.

[0162] Alternatively, in the terminal devices 100 (100A, 100B, 100C, 100D), 100-1, 100-2, 100-3, ..., 100-n, the global model receiving unit 110, the global model evaluation unit 120 (120D), the local model updating unit 130 (130B, 130C, 130D), part of the local model holding unit 140, the data collecting unit 150, the local model learning unit 160, the local model information output unit 170 (170C, 170D), the individualization status calculation unit 180, and some of the functions of the terminal-side control unit (not shown) may be realized by the processor 10001 and the memory 10002, and the remaining functions may be realized by the processing circuit 20001. Similarly, in the server 300 (300C, 300D), some of the functions of the local model information receiving unit 310 (310C, 310D), the global model integration unit 320 (320C, 320D), the global model distribution unit 330 (330C, 330D), and the server-side control unit (not shown) may be realized by the processor 10001 and memory 10002, and the remaining functions may be realized by the processing circuit 20001.

[0163] It should be noted that, within the scope of this disclosure, the embodiments may be freely combined, any component of each embodiment may be modified, or any component of each embodiment may be omitted.

[0164] The present disclosure can improve the accuracy of learning models for clients compared to conventional federated learning, and is therefore suitable for use in, for example, federated learning systems, terminal devices, servers, model update methods, etc. that utilize federated learning mechanisms.

[0165] 1 (1A, 1B, 1C, 1D) Federated learning system, 100 (100A, 100B, 100C, 100D), 100-1, 100-2, 100-3, ..., 100-n Terminal device, 110 Global model reception unit, 120 (120D) Global model evaluation unit, 130 (130B, 130C, 130D) Local model update unit, 140 Local model storage unit, 150 Data collection unit, 160 Local model learning unit, 170 (170C, 170D) Local model information output unit, 180 Individualization status calculation unit, 200 (200-1, 200-2, 200-3, ..., 200-n) Information source device, 300 (300C, 300D) Server, 310 (310C, 310D) Local model information receiving unit, 320 (320C, 320D) global model integration unit, 330 (330C, 330D) global model distribution unit, 2000 model (local model), 2100 input layer, 2400 final layer (nodes, parameters), 2500 output layer, 4000 model (local model), 4100 input layer, 4400 final layer (nodes, parameters), 4500 output layer, 4600 final layer (nodes, parameters) of class C, 10001 processor, 10002 memory, 10003 input / output interface, 10004 communication circuit, 20001 processing circuit, 20002 memory, 20003 input / output interface, 20004 communication circuit.

Claims

1. A terminal device comprising: a global model acceptance unit that accepts an updated global model by integrating multiple trained local models, each of which is a local model trained in a different terminal device; a global model evaluation unit that determines whether to adopt the global model based on evaluation results of the global model accepted by the global model acceptance unit and a currently held local model; and a local model update unit that determines a method for updating the local model based on the evaluation results by the global model evaluation unit, and updates the held local model in accordance with the determined update method.

2. The terminal device according to claim 1, characterized in that it comprises: a local model holding unit that holds the local model updated by the local model update unit; a data collection unit that collects learning data for training the local model held in the local model holding unit; and a local model learning unit that learns and constructs the local model held in the local model holding unit based on the learning data collected by the data collection unit.

3. The terminal device of claim 1 or claim 2, characterized in that the local model update unit, when the evaluation result by the global model evaluation unit indicates that the performance of the global model accepted by the global model acceptance unit is lower than the performance of the currently held local model, determines an update method for updating the held local model by combining a portion of the held local model with a portion of the global model.

4. The terminal device of claim 1 or claim 2, characterized in that the local model update unit determines the update method by gradually changing the combination of a portion of the held local model and a portion of the global model when the evaluation result by the global model evaluation unit indicates that the performance of the global model accepted by the global model acceptance unit is lower than the performance of the currently held local model.

5. The terminal device described in claim 2, characterized in that it comprises: an individualization status calculation unit that calculates an individualization status indicating the degree to which a local model updated based on the update method determined by the local model update unit is individualized relative to the global model; and a local model information output unit that outputs the individualization status to the outside of the device together with a learned local model, which is a local model learned by the local model learning unit.

6. A terminal device according to any one of claims 1 to 5, characterized in that the global model evaluation unit performs evaluation for each class that can be output using the global model accepted by the global model acceptance unit and the currently held local model, and the local model update unit determines an update method for each class using the evaluation results for each class by the global model evaluation unit.

7. A federated learning system that performs federated learning using a local model on a terminal device side and a global model on a server side, wherein the terminal device comprises: a global model reception unit that receives an updated global model by integrating multiple trained models, which are multiple local models trained on different terminal devices; a global model evaluation unit that determines whether to adopt the global model based on evaluation results of the global model received by the global model reception unit and the local model currently held; a local model update unit that determines an update method for the local model based on the evaluation results by the global model evaluation unit and updates the held local model in accordance with the determined update method; a local model holding unit that holds the local model updated by the local model update unit; a data collection unit that collects learning data for training the local model held in the local model holding unit; a local model learning unit that learns and constructs the local model held in the local model holding unit based on the learning data collected by the data collection unit; and an individualization status calculation unit that calculates an individualization status that indicates the degree to which the local model updated based on the update method determined by the local model update unit is individualized with respect to the global model. a local model information output unit that outputs a trained local model, which is a local model trained by the local model learning unit, and the individualization status to the server outside the device, wherein the server comprises: a local model information receiving unit that receives the trained local model trained on the terminal device and the individualization status; a global model integration unit that updates a global model by integrating multiple trained local models, which are local models trained on different terminal devices, in accordance with the individualization status of each trained local model; and a global model distribution unit that outputs the global model updated by the global model integration unit to the terminal device.

8. A server comprising: a local model information receiving unit that receives a local model learned in a terminal device and an individualization status indicating the degree to which the local model is individualized relative to a global model; a global model integration unit that updates a global model by integrating multiple learned local models, each of which is a local model learned in a different terminal device, according to the individualization status of each learned local model; and a global model distribution unit that outputs the global model updated by the global model integration unit to the terminal device.

9. A model updating method for updating a local model used in a terminal device based on a global model received in the terminal device from outside the device, wherein a global model receiving unit of the terminal device receives an updated global model by integrating multiple trained local models, each of which is a local model trained in a different terminal device; a global model evaluation unit of the terminal device evaluates the global model received by the global model receiving unit and a currently held local model to determine whether to adopt the global model; and a local model updating unit of the terminal device determines a method for updating the local model based on the evaluation result by the global model evaluation unit, and updates the currently held local model in accordance with the determined update method.

Images