WO2019171443A1

WO2019171443A1 - Machine learning-trained model update system, edge device, machine learning-trained model update method, and program

Info

Publication number: WO2019171443A1
Application number: PCT/JP2018/008419
Authority: WO
Inventors: 篤古城; 将仁谷口
Original assignee: 株式会社ウフル
Priority date: 2018-03-05
Filing date: 2018-03-05
Publication date: 2019-09-12

Abstract

In the field of Internet of Things (IoT), the present invention provides a mechanism whereby a machine learning-trained model installed on an edge device not connected to the Internet is updated to a machine learning-trained model provided via the Internet. When an edge device (130), which is not directly connected to the Internet, is in a state where the edge device (130) is connected to the Internet via a communication device (120) connected to the Internet, a machine learning-trained model update system (100) updates a machine learning-trained model installed on the edge device (130) to a machine learning-trained model provided via the Internet.

Description

Machine learning model update system, edge device, machine learning model update method, and program

The present invention relates to a technique for updating a machine-learned model installed in an edge device, and is used in the field of IoT (Internet of Things).

In recent years, with the advancement of AI (artificial intelligence) technology, renewed technology for learned models has attracted attention. For example, a technique for updating a model of a machine learning system is provided (Patent Document 1).

JP 2017-120647 A

In the system described in Patent Document 1, a machine learning system model is trained in a single device, but an edge device that does not have the processing capability to train a machine-learned model is machine-learned by itself. Unable to train the system model. In order to update the machine-learned model installed on such an edge device, for example, a server may update the machine-learned model and provide the updated model to the edge device via the Internet. . However, some medical devices and devices installed in factories are not connected to the Internet from the viewpoint of protecting personal information and ensuring security. For such devices, the installed machine-learned model cannot be updated by the method described above.
An object of the present invention is to provide a mechanism for updating a machine-learned model installed in an edge device that is not connected to the Internet to a machine-learned model provided via the Internet.

The present invention provides a machine-learned model installed in an edge device that is not connected to the Internet when the edge device is connected to the Internet via a communication device connected to the Internet. A machine-learned model update system for updating to a new machine-learned model provided via the Internet is provided.

Furthermore, the present invention is an edge device that is not connected to the Internet, and the machine-learned model according to claim 1 when the device is connected to the Internet via a communication device connected to the Internet. An edge device is provided that acquires the new machine-learned model from the update system via the Internet and replaces the installed machine-learned model with the new machine-learned model.

Furthermore, the present invention provides a machine-learned model that is installed on an edge device that is not connected to the Internet when the edge device is connected to the Internet via a communication device connected to the Internet. Is updated to a new machine-learned model provided via the Internet.

Furthermore, the present invention relates to a machine installed in an edge device that is not connected to the Internet when the edge device that is not connected to the Internet is connected to the Internet via a communication device connected to the Internet. A program for executing a step of updating a learned model to a new machine-learned model provided via the Internet is provided.

According to the present invention, a machine-learned model installed on an edge device that is not connected to the Internet can be updated to a machine-learned model provided via the Internet.

It is a figure which shows an example of the machine learning completed model update system 100 which concerns on embodiment. 2 is a diagram illustrating an example of a hardware configuration of a server 110. FIG. 2 is a diagram illustrating an example of a functional configuration of a server 110. FIG. 2 is a diagram illustrating an example of a hardware configuration of an edge device 130. FIG. 3 is a diagram illustrating an example of a functional configuration of an edge device 130. FIG. 5 is a sequence chart showing an example of an operation of the machine learning completed model update system 100. It is a figure which shows an example of the posting information 410. FIG. It is a figure which shows an example of the machine learning completed model update system 200 which concerns on a modification. It is a figure which shows an example of a function structure of the server 110 which concerns on a modification.

1. Configuration FIG. 1 is a diagram illustrating an example of a machine-learned model update system 100 according to the present embodiment. The machine-learned model update system 100 is a system that updates a machine-learned model installed in an edge device 130 that is not connected to the Internet. This “installation” means that the computer is stored and ready for use.

The machine learning completed model update system 100 includes a server 110, a communication device 120, and a plurality of edge devices 130. In addition, the number of each apparatus shown in FIG. 1 is an illustration, and is not limited to this. The server 110 and the communication device 120 are connected via a communication network 140. The communication network 140 includes the Internet. The plurality of edge devices 130 are not connected to the communication network 140.

FIG. 2 is a diagram illustrating an example of a hardware configuration of the server 110. The server 110 is a device that generates a machine-learned model using learning data and provides the generated machine-learned model to the edge device 130. The server 110 is a computer including a processor 111, a memory 112, a storage 113, and a communication unit 114, for example. These devices are connected via a bus 115.

The processor 111 executes various processes by reading the program into the memory 112 and executing it. The processor 111 may be configured by, for example, a CPU (Central Processing 例えば Unit). The memory 112 stores a program executed by the processor 111. The memory 112 may be configured by, for example, a ROM (Read Only Memory) or a RAM (Random Access Memory). The storage 113 stores various data and programs. The storage 113 may be configured by, for example, a hard disk drive or a flash memory. The storage 113 stores a server program for realizing the function of the server 110. The communication unit 114 is a communication interface connected to the communication network 140. The communication unit 114 performs data communication via the communication network 140.

FIG. 3 is a diagram illustrating an example of a functional configuration of the server 110. The server 110 includes a generation unit 211, a posting unit 212, an edge device authentication unit 213, and a control unit 214. These functions are realized by the processor 111 performing calculations or controlling communication by the communication unit 114 in cooperation with the server program stored in the memory 112 or the storage 113 and the processor 111 that executes the server program. Is done.

The generation unit 211 generates a new machine learned model corresponding to the machine learned model installed in the edge device 130 (hereinafter referred to as “new machine learned model”). The new machine-learned model is generated by machine learning of learning data using a learning device, for example. The new machine-learned model may be an improved model of at least a part of the machine-learned model, for example, a machine-learned model that is updated to improve accuracy. Alternatively, the new machine-learned model may be a modified machine-learned model so that a new purpose can be achieved. The learning data is data used for machine learning. For the learning data, for example, sensor data output from various sensors connected to the Internet is used.

The posting unit 212 posts to the edge device 130 that there is a new machine learned model. This “posting” means presenting information. This posting may be realized, for example, by storing information indicating that there is a new machine learned model in the storage 113.

The edge device authentication means 213 authenticates the edge device 130 using the authentication key and ID (Identification) of the edge device 130. The authentication key is information used for authentication of the edge device 130. The ID is information for identifying the edge device 130. The authentication key and ID of the edge device 130 are stored in the storage 113 in advance, for example.

The control means 214 controls the edge device 130 to acquire a new machine learned model and update the machine learned model. This “update” means replacing a machine-learned model with a new one. This control may be realized by, for example, acquiring a new machine learned model and transmitting a control signal for updating the machine learned model.

Referring back to FIG. 1, the communication device 120 is a device having a communication function such as a smartphone, a tablet terminal, or a personal computer. The communication device 120 is carried by the user and moves. The communication device 120 is connected to the server 110 via the communication network 140. The communication device 120 is connected to the edge device 130 via a communication network 150 different from the communication network 140. The communication network 150 is, for example, a wireless network. However, the communication network 150 is not limited to a wireless network, and may be a wired network. When the communication device 120 is connected to the edge device 130, the communication device 120 relays data exchange performed between the edge device 130 and the server 110. At this time, the communication device 120 communicates with the edge device 130 according to a predetermined communication standard. The predetermined communication standard is a short-range wireless communication standard such as Wi-Fi (registered trademark), Bluetooth (registered trademark), ZigBee (registered trademark), or the like. However, the predetermined communication standard is not limited to the short-range wireless communication standard, and may be a long-range wireless communication standard or a wired communication standard.

FIG. 4 is a diagram illustrating an example of a hardware configuration of the edge device 130. The edge device 130 is not connected to the Internet, and performs a predetermined process using a machine-learned model. The edge device 130 is a device such as a dialysis machine not connected to the Internet, an MRI (Magnetic Resonance Imaging) device, a CT (computed Tomography) scan, a medical device such as an electronic medical record system, and various devices installed in a factory. . However, the edge device 130 is not limited to these devices, and may be any device as long as it performs processing using a machine-learned model.

The edge device 130 is a computer including a processor 131, a memory 132, a storage 133, and a communication unit 134, for example. These devices are connected via a bus 135. Since the processor 131, the memory 132, the storage 133, the communication unit 134, and the bus 135 are the same as the processor 111, the memory 112, the storage 113, the communication unit 114, and the bus 115 described above, description thereof is omitted. However, the communication unit 134 is a communication interface connected to the communication network 150. When connected to the communication network 150, the communication unit 134 performs data communication via the communication network 150. In addition, the storage 133 stores a client program for realizing the function of the edge device 130 and a machine-learned model used for processing executed in the edge device 130.

FIG. 5 is a diagram illustrating an example of a functional configuration of the edge device 130. The edge device 130 includes a connection unit 231, a server authentication unit 232, a confirmation unit 233, an acquisition unit 234, a replacement unit 235, and a processing unit 236. These functions are realized by the processor 131 performing calculations or controlling communication by the communication unit 134 in cooperation with the client program stored in the memory 132 or the storage 133 and the processor 131 that executes the program program. Is done.

The connection unit 231 establishes a connection with the communication device 120 via the communication network 150. The connection is established when the edge device 130 enters the communication range of the communication device 120 by moving the communication device 120, for example.

The server authentication unit 232 authenticates the server 110 using the authentication key of the server 110. The authentication key is information used for authentication of the server 110. The authentication key is stored in the storage 133 in advance, for example.

The confirmation unit 233 is connected to the communication network 140 via the communication device 120, and based on the posted content of the server 110, the new machine learning corresponding to the machine learned model installed in the edge device 130 Check if there is a finished model. This confirmation may be performed by accessing the posting information 410 stored in the storage 113 of the server 110, for example.

The acquisition unit 234 acquires a new machine learned model from the server 110. More specifically, the acquisition of the new machine-learned model is performed by pull distribution. That is, the new machine learned model is transmitted in response to a request from the acquisition unit 234. At this time, the new machine learned model may be transmitted using an encrypted communication method. This encrypted communication method refers to a communication method in which data is encrypted for communication. As the encrypted communication method, a known encrypted communication method such as SSL (Secure Sockets Layer) or TLS (Transport Layer Security) may be used.

The replacement unit 235 obtains the new machine learning obtained by the obtaining unit 234 according to the control by the control unit 214 of the server 110 and the machine learned model stored in the storage 133 (hereinafter also referred to as “current machine learned model”). Replace with a completed model. This replacement means making the new machine learned model usable in place of the current machine learned model. The replacement may be realized, for example, by storing the new machine learned model in the storage 133 and setting the new machine learned model to be used instead of the current machine learned model. At this time, the current machine learned model stored in the storage 133 may be deleted or stored as it is.

The processing unit 236 executes a predetermined process using the new machine learned model replaced by the replacement unit 235. The predetermined process may be different for each edge device 130. For example, when the edge device 130 is an MRI apparatus, the predetermined process may be a process of capturing a tomographic image inside the body. On the other hand, when the edge device 130 is an electronic medical record system, the predetermined process may be a process of creating an electronic medical record.

2. Operation FIG. 6 is a sequence chart illustrating an example of the operation of the machine-learned model update system 100. Here, it is assumed that the communication device 120 and the edge device 130 correspond to Bluetooth (registered trademark) and are paired in advance. In this case, communication between the communication device 120 and the edge device 130 is performed according to Bluetooth (registered trademark).

The server 110 generates a new machine-learned model by causing the generation unit 211 to machine-learn the learning data using a learning device (step 301). The new machine-learned model may be, for example, a model obtained by improving the accuracy of the current machine-learned model, or may be a model obtained by modifying the current machine-learned model so that a new analysis can be performed. This new machine learned model is stored in the storage 113.

The server 110 causes the bulletin board 212 to store bulletin information 410 indicating that there is a new machine learned model in the storage 113 (step 302).

FIG. 7 is a diagram illustrating an example of the bulletin information 410. Here, it is assumed that a new machine learned model for the edge device 130 with the device ID “001” is generated. In this case, the device ID “001”, information “Yes” indicating that there is a new machine learned model, and the acquisition destination of this new machine learned model, that is, the new machine learned model is stored in the storage 113. An address “http://www.example.com/M001” indicating the location is stored in association with each other.

When updating the machine-learned model of the edge device 130 whose device ID is “001”, the user approaches the edge device 130 with the communication device 120. When the edge device 130 enters the communication range of the communication device 120, the edge device 130 establishes a connection with the communication device 120 via the communication network 150 by the connection means 231 (step 303). As a result, the edge device 130 is connected to the communication network 140 via the communication device 120. The state of being connected to the communication network 140 via the communication device 120 is not directly connected to the communication network 140, but communication via the communication network 140 is possible when the communication device 120 relays. State.

When the edge device 130 detects that it is connected to the communication network 140 via the communication device 120, the server authentication means 232 authenticates the server 110 using the authentication key of the server 110 (step 304). Specifically, the authentication key of the server 110 is stored in advance in the storage 133, for example. The edge device 130 transmits an authentication key (hereinafter referred to as “target authentication key”) stored in the storage 133 to the communication device 120. When receiving the target authentication key from the edge device 130, the communication device 120 transmits the target authentication key to the server 110. When the target authentication key is the authentication key of the server 110, the server 110 transmits a response indicating that to the communication device 120. When receiving the response from the server 110, the communication device 120 transmits this response to the edge device 130. When a response indicating that the target authentication key is the authentication key of the server 110 is received, the server authentication unit 232 authenticates the server 110. On the other hand, when a response indicating that the target authentication key is not the authentication key of the server 110 is received, the server 110 is not authenticated and the subsequent processing is not performed. In the following description, transmitting data to the communication device 120 and transferring the data to the server 110 or the edge device 130 is referred to as “transmitting via the communication device 120”.

The server 110 authenticates the edge device 130 by the edge device authentication means 213 using the authentication key and ID of the edge device 130 (step 305). Specifically, the authentication key and ID of the edge device 130 are stored in advance in the storage 113, for example. The server 110 transmits the authentication key and ID stored in the storage 113 (hereinafter referred to as “target authentication key and ID”) to the edge device 130 via the communication device 120. When the target authentication key and ID are the authentication key and ID of the edge device 130, the edge device 130 transmits a response indicating that to the server 110 via the communication device 120. When the response indicating that the target authentication key and ID are the authentication key and ID of the edge device 130 is received, the edge device authentication unit 213 authenticates the edge device 130. On the other hand, when a response indicating that the target authentication key and ID are not the authentication key and ID of the edge device 130 is received, the edge device 130 is not authenticated and the subsequent processing is not performed.

When the server 110 and the edge device 130 are mutually authenticated, the edge device 130 accesses the bulletin information 410 stored in the storage 113 via the communication device 120 by the confirmation unit 233 (step 306). Subsequently, the edge device 130 determines whether or not there is a new machine learned model corresponding to the machine learned model installed in the edge device 130 based on the posting information 410 (step 233). 307). In the example illustrated in FIG. 7, the bulletin information 410 stores information “Yes” indicating that there is a new machine learned model in association with the device ID “001”. In this case, it is determined that there is a new machine learned model generated for the edge device 130 (YES in step 307), and the process proceeds to the next step. On the other hand, when information “none” indicating that there is no new machine learned model is stored in association with the device ID “001”, there is no new machine learned model generated for the edge device 130. Is determined (NO at step 307), and the process ends.

In accordance with step 306 described above, the server 110 acquires a new machine learned model using a predetermined encrypted communication method, and makes the current machine learned model stored in the storage 133 a new machine learned model. A control signal for replacement is transmitted from the control unit 214 to the edge device 130 whose device ID is “001” via the communication device 120 (step 308).

When receiving the control signal, the edge device 130 uses the predetermined encrypted communication method in accordance with the control signal, and “http://www.example.com/M001” included in the posting information 410 shown in FIG. The new machine learned model is acquired by the acquiring unit 234 from the location indicated by the address (step 309). Specifically, a request for acquiring a new machine learned model is transmitted to the server 110 via the communication device 120, and the new machine learned model transmitted from the server 110 via the communication device 120 in response to the request is transmitted. Received by the acquisition means 234.

The edge device 130 replaces the current machine-learned model stored in the storage 133 with the new machine-learned model by the replacing unit 235 according to the received control signal (step 310). Specifically, the new machine learned model is stored in the storage 133 instead of the current machine learned model, and the new machine learned model is used instead of the current machine learned model.

After the machine-learned model is updated in this way, the edge device 130 executes a predetermined process by the processing unit 236 using the new machine-learned model (step 311).

In the above-described embodiment, “storage 113”, “ID of edge device 130”, and “new machine learned model” are “storage means”, “identification information”, and “new machine” according to the present invention, respectively. Used as a “learned model”.

According to the embodiment described above, the machine-learned model installed in the edge device 130 that is not directly connected to the Internet by communicating with the server 110 via the communication device 120 connected to the Internet, It is possible to update to a new machine learned model provided from the server 110. In addition, the generation of a new machine-learned model requires a high processing capability. However, in the above-described embodiment, since the server 110 generates a new machine-learned model, the edge device 130 has a high processing capability. You don't have to.

Furthermore, in the above-described embodiment, the bulletin information 410 and the new machine learned model are transmitted in response to pull distribution from the edge device 130, that is, in response to a request from the edge device 130. If the server 110 adopts a push distribution, that is, a configuration in which the server 110 transmits the data without requesting it from the edge device 130, the edge device 130 impersonates the server 110 as well as the server 110. It is also possible to receive information transmitted by push-side distribution from other devices. In this case, there is a risk that damage is caused by information transmitted by another malicious device. In contrast, as in the above-described embodiment, when the data is pulled from the server 110, that is, when the data is transmitted from the server 110 in response to a request from the edge device 130, such a risk is reduced, so that information security is high. Become.

Furthermore, according to the above-described embodiment, since the server 110 is authenticated by the edge device 130, it is possible to prevent impersonation of the server 110 by another malicious device. Furthermore, according to the above-described embodiment, since authentication of the edge device 130 is performed in the server 110, it is possible to prevent spoofing of the edge device 130 by another malicious device. Furthermore, according to the above-described embodiment, since the new machine learned model is acquired using the encrypted communication method, it is possible to prevent the new machine learned model from being acquired by a malicious third party.

3. The present invention is not limited to the above-described embodiment. Various modifications may be made to the above-described embodiment. Moreover, the following modifications may be implemented in combination.

In the embodiment described above, when a plurality of edge devices 130 perform similar processing, the same machine-learned model may be installed in these edge devices 130. In this case, one edge device 130 included in the plurality of edge devices 130 may distribute the new machine learned model to other edge devices 130.

FIG. 8 is a diagram illustrating an example of the machine learning completed model update system 200 according to the modification. Similar to the machine learning completed model update system 100 described above, the machine learned model update system 200 includes a server 110, a communication device 120, and a plurality of edge devices 130. However, the plurality of edge devices 130 are connected via the communication network 160. The communication network 160 is a closed network such as a LAN (Local Area Network) or a dedicated line, and is not connected to the communication network 140. Each of the plurality of edge devices 130 performs data communication with other edge devices 130 by the communication unit 134. The same machine-learned model is stored in advance in the storage 133 of the plurality of edge devices 130.

In this case, the one edge device 130 acquires the new machine learned model from the server 110 as in the above-described embodiment. In addition, the edge device 130 transmits a copy of the acquired new machine-learned model to another edge device 130 via the communication network 160. This duplicate is generated, for example, by copying a new machine learned model. Note that when a copy of the new machine-learned model is transmitted to another edge device 130, the encrypted communication method may not be used. When receiving the new machine learned model, each of the other edge devices 130 replaces the current machine learned model stored in the storage 133 with the new machine learned model, similarly to step 310 described above. In another example, when a management apparatus is connected to the communication network 160, the management apparatus may perform processing of one edge device 130 instead of the one edge device 130. According to this modification, the machine-learned models installed in the plurality of edge devices 130 can be updated all at once.

In the embodiment described above, the client program may be stored in a device externally attached to the edge device 130. In this case, even if the edge device 130 has no client program installed, the processing described in the above-described embodiment can be performed by externally attaching this device.

In the above-described embodiment, the server 110 evaluates the current machine-learned model and the new machine-learned model used in the edge device 130, and when the current machine-learned model has higher accuracy. The new machine learned model may not be provided to the edge device 130. In this case, the current machine-learned model used in the edge device 130 is stored in the storage 113 of the server 110 in addition to the new machine-learned model.

FIG. 9 is a diagram illustrating an example of a functional configuration of the server 110 according to the modification. The server 110 includes an evaluation unit 215 in addition to the functions shown in FIG. The evaluation means 215 evaluates the current new machine learned model and the new machine learned model. This evaluation may be performed, for example, by inputting learning data and obtaining the prediction accuracy of these machine-learned models. For this evaluation, a well-known evaluation index such as a ROC (receiver operating characteristic curve) curve or AUC (Area under an ROC curve) may be used. In this case, the posting unit 212 determines whether the accuracy of the new machine learned model is higher than the accuracy of the current machine learned model based on the evaluation result of the evaluation unit 215. When the accuracy of the new machine-learned model is higher than the accuracy of the current machine-learned model, the posting unit 212 posts information indicating that there is a new machine-learned model, as in the above-described embodiment. On the other hand, when the accuracy of the new machine learned model is lower than the accuracy of the current machine learned model, the posting unit 212 does not post information indicating that there is a new machine learned model. In this case, since the determination in Step 307 described above is NO, the edge device 130 does not update the machine-learned model. A machine-learned model has a problem of overlearning, and a machine-learned model with advanced learning is not necessarily superior. According to this modification, when the new machine learned model is inferior to the current machine learned model, it is possible to prevent the current machine learned model from being updated.

In the embodiment described above, the machine learning algorithm generates a model based on the result of machine learning from given data, for example, on a computer, and inputs new input data to the model. A so-called supervised learning algorithm for outputting an event predicted from the input data may be used. However, the machine learning algorithm is not limited to a so-called supervised learning algorithm, and may be an algorithm for machine learning such as unsupervised learning, semi-supervised learning, reinforcement learning, and expression learning. Further, the algorithm for machine learning may include other learning algorithms such as data mining and deep learning. Note that these learning algorithms use various techniques or techniques such as decision tree learning, correlation rule learning, neural network, genetic programming, inductive logic programming, support vector machine, clustering, and Bayesian network. included. In short, any machine learning algorithm may be used as long as it is processed together with some data provided by the data provider and outputs information desired by the user as a result of the processing.

In the embodiment described above, the update of the machine-learned model is not limited to an update that completely replaces the current machine-learned model, and may be an update that replaces only the difference from the current machine-learned model. In this case, only the difference of the new machine learned model may be provided to the edge device 130.

In the above-described embodiment, the location where the new machine learned model is stored is not limited to the storage 113 of the server 110. For example, when a storage device is connected to the communication network 140, the new machine learned model may be stored in this storage device instead of the storage 113. In this case, the posting information 410 includes an address indicating a location in the storage device. Then, the edge device 130 acquires a new machine learned model from the storage device based on this address.

In the embodiment described above, the posting information 410 may be updated in response to the acquisition of the new machine learned model by the edge device 130. The determination whether or not a new machine learned model has been acquired may be performed based on whether or not information indicating that a new machine learned model has been acquired from the edge device 130, for example. For example, when the new machine learned model is acquired by the edge device 130 with the device ID “001”, the new machine learned model stored in association with the device ID “001” in the posting information 410 illustrated in FIG. Information indicating presence / absence is changed from “present” to “none”. In addition, it is changed to “-” indicating that the acquisition destination of the new machine learned model is not applicable. Thereby, it is possible to prevent the edge device 130 from trying to acquire the acquired new machine learned model again.

In the above-described embodiment, the machine-learned model does not necessarily have to be installed in advance in the edge device 130. For example, a machine learned model provided from the server 110 may be installed in the edge device 130.

The subject that realizes the functions of the machine learning completed

model update system

100 or 200 is not limited to the example described in the above-described embodiment. For example, the edge device 130 or an external device may have a part of the function of the server 110. For example, an external device used by a provider such as an AI (Artificial Intelligent) vendor may have the generation unit 211 instead of the server 110. In this case, the external device generates a new machine learned model using the learning data, and transmits the generated new machine learned model to the server 110. As described above, the new machine learned model may be generated on the provider side and registered in the server 110. Further, the server 110 or an external device may have a part of the function of the edge device 130.

The steps of processing performed in the machine learning completed model update system 100 are not limited to the example described in the above embodiment. The steps of this process may be interchanged as long as there is no contradiction. Further, the present invention may be provided as a machine-learned model update method performed in the machine-learned

model update system

100 or 200.

The present invention may be provided as a program executed in the server 110, the communication device 120, or the edge device 130. This program may be downloaded via a communication line such as the Internet, or a computer such as a magnetic recording medium (magnetic tape, magnetic disk, etc.), an optical recording medium (optical disk, etc.), a magneto-optical recording medium, or a semiconductor memory. May be provided in a state of being recorded on a readable recording medium.

100, 200: Model update system, 110: Server, 120: Communication device, 130: Edge device, 211: Generation means, 212: Posting means, 213: Edge device authentication means, 214: Control means, 215: Evaluation means, 231 : Connection means, 232: server authentication means, 233: confirmation means, 234: acquisition means, 235: replacement means, 236: processing means

Claims

When an edge device not connected to the Internet is connected to the Internet via a communication device connected to the Internet, a machine-learned model installed on the edge device is transferred via the Internet. A machine-learned model update system that updates to the new machine-learned model provided.
The machine-learned model update system according to claim 1, further comprising server authentication means for authenticating the server using an authentication key of the server providing the new machine-learned model.
The machine-learned model update system according to claim 1, further comprising an edge device authentication unit that authenticates the edge device using an authentication key of the edge device.
The machine-learned model update system according to claim 3, wherein the edge device authentication unit authenticates the edge device using the authentication key and identification information of the edge device.
5. The control unit according to claim 1, further comprising: a control unit configured to cause the edge device to acquire the new machine-learned model and update the machine-learned model to the acquired new machine-learned model. The machine learning completed model update system according to claim 1.
The machine learning model update system according to claim 5, wherein the control unit controls the edge device to acquire the new machine learned model using an encrypted communication method.
The edge device is connected to another edge device not connected to the Internet via a communication network different from the Internet, and sends a copy of the new machine-learned model to the other edge device. The machine-learned model update system according to any one of claims 1 to 6, wherein the machine-learned model installed in the other edge device is updated to the new machine-learned model.
Storage means for storing the generated new machine-learned model when a new machine-learned model corresponding to the machine-learned model is generated by machine learning of learning data using a learning device; ,
Posting means for posting on the edge device information indicating that there is a new machine-learned model and an address indicating a location in the storage means in which the new machine-learned model is stored;
The edge device is
Connection means for establishing a connection with the communication device via a communication network different from the Internet;
After the connection is established, if it is determined that there is the new machine-learned model based on the posted information, the new address is sent from the location indicated by the posted address via the communication device. An acquisition means for acquiring a machine-learned model;
Replacement means for replacing the machine-learned model with the acquired new machine-learned model;
The machine-learned model update system according to any one of claims 1 to 7, further comprising a processing unit that executes a predetermined process using the new machine-learned model.
An evaluation means for evaluating the machine-learned model and the new machine-learned model;
If the accuracy of the new machine-learned model is evaluated to be lower than the accuracy of the machine-learned model, the posting unit does not post information indicating that there is the new machine-learned model. Item 9. The machine-learned model update system according to Item 8.
An edge device that is not connected to the Internet,
The new machine-learned model is acquired via the Internet from the machine-learned model update system according to claim 1 when connected to the Internet via a communication device connected to the Internet. An edge device that replaces an installed machine-learned model with the new machine-learned model.
When an edge device not connected to the Internet is connected to the Internet via a communication device connected to the Internet, a machine-learned model installed on the edge device is transferred via the Internet. A machine-learned model update method that updates to a new machine-learned model that is provided.
On the computer,
When an edge device not connected to the Internet is connected to the Internet via a communication device connected to the Internet, a machine-learned model installed on the edge device is transferred via the Internet. A program to execute the step of updating to a new machine-learned model provided.