Learning devices, learning systems, learning methods, and programs

Abstract

Provided are a learning device, a learning system, a learning method, and a program capable of building a global model in a case where a plurality of organization networks are not constantly connected. A learning device (1) comprises: a communication establishment unit (11) that establishes secure communications with information terminals (2) arranged in networks (N) of the respective organizations; an acquisition unit (12) that acquires a data set for each organization from the information terminals (2) by using secure communications; a learning unit (13) that causes a local model to learn the data set; and an integration unit (14) that integrates the plurality of local models which have learned the plurality of data sets.

Description

Technical Field

[0001] The present disclosure relates to a learning device, a learning system, a learning method, and a computer-readable medium.

Background Art

[0002] Patent Document 1 discloses a technique for performing machine learning to construct an AI (Artificial Intelligence) model (also referred to as a local model) individualized for a user.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] It is known that a better-performing AI model (also referred to as a global model) can be constructed by integrating a plurality of local AI models. By collecting user data by the server, the server can construct local models and global models.

[0005] When the user is an organization, it is necessary to collect data owned by each organization, so it is desirable to construct a network connecting a plurality of organizations. However, there has been a problem that it is difficult to construct a network connecting a plurality of organizations with different views on security.

[0006] Therefore, one of the objects to be achieved by the embodiments disclosed in this specification is to provide a learning device, a learning system, a learning method, and a computer-readable medium capable of constructing a global model even when the networks of a plurality of organizations are not always connected.

Means for Solving the Problems

[0007] The learning device relating to the first aspect of this disclosure is A means for establishing secure communication with information terminals located on the network of each organization, An acquisition means that uses the aforementioned secure communication to acquire a dataset for each organization from the information terminal, A learning method for training a local model using the aforementioned dataset, An integration method for combining multiple local models that have been trained on multiple datasets. It is equipped with.

[0008] The computing system relating to the second aspect of this disclosure is Information terminals located on each organization's network, Learning device, A learning system equipped with, The learning device is Secure communication is established with the aforementioned information terminal, Using the aforementioned secure communication, obtain the organization-specific dataset from the information terminal. The aforementioned dataset is used to train a local model. Integrate multiple local models that have been trained on multiple datasets.

[0009] The calculation method relating to the third aspect of this disclosure, Computers Secure communication is established between information terminals located on each organization's network. Using the aforementioned secure communication, obtain the organization-specific dataset from the information terminal. The aforementioned dataset is used to train a local model. Integrate multiple local models that have been trained on multiple datasets.

[0010] In the non-temporary computer-readable media relating to the fourth aspect of this disclosure, On the computer, A process for establishing secure communication with an information terminal arranged in a network of each organization, and A process for acquiring a data set for each organization from the information terminal using the secure communication, and A process for training the data set on a local model, and A process for integrating a plurality of local models trained with a plurality of data sets and A program for executing the above is stored.

Effect of the Invention

[0011] According to the present disclosure, it is possible to provide a learning device, a learning system, a learning method, and a computer-readable medium capable of constructing a global model even when the networks of a plurality of organizations are not always connected.

Brief Description of the Drawings

[0012] [Figure 1] It is a block diagram showing the configuration of a learning device according to Embodiment 1. [Figure 2] It is a block diagram showing the configuration of a learning system according to Embodiment 2. [Figure 3] It is a block diagram showing the configuration of a learning device according to Embodiment 2. [Figure 4] It is a flowchart showing the flow of an operation for generating a local model. [Figure 5] It is a block diagram showing the configuration of a learning system according to Embodiment 2.

Modes for Carrying Out the Invention

[0013] <Embodiment 1> FIG. 1 is a block diagram showing the configuration of the learning device 1 according to Embodiment 1. The learning device 1 includes a communication establishment unit 11, an acquisition unit 12, a learning unit 13, and an integration unit 14. The learning device 1 is connected to a public network (not shown). Each organization's network is connected to the public network. An information terminal (not shown) is arranged in each organization's network. The information terminal is a repository in which a dataset owned by each organization is stored.

[0014] The communication establishment unit 11 establishes secure communication with the information terminal arranged in each organization's network. The communication establishment unit 11 may establish secure communication at a predetermined timing. The communication establishment unit 11 may establish secure communication based on the progress of learning of the local model described later.

[0015] The communication establishment unit 11, for example, connects the learning device 1 to each organization's network via a VPN (Virtual Private Network). In this case, the communication between the learning device 1 and the information terminal is kept confidential by encryption or encapsulation. That is, secure communication is established between the learning device 1 and the information terminal.

[0016] Note that the communication establishment unit 11 may establish secure communication using a technology other than VPN. The communication establishment unit 11 may control communication by a protocol including encryption (e.g., SSL / TLS, SSH (Secure Shell), FTPS (File Transfer Protocol over SSL / TLS)).

[0017] The acquisition unit 12 acquires a dataset for each organization from the information terminal using secure communication.

[0018] The learning unit 13 causes the local model to learn from the dataset.

[0019] The integration unit 14 integrates a plurality of local models that have learned from a plurality of datasets.

[0020] The learning device 1 includes a processor, memory, and storage device (not shown in the diagram). The storage device stores a computer program that implements the learning method according to this embodiment. The processor loads the computer program from the storage device into the memory and executes the computer program. In this way, the processor realizes the functions of the communication establishment unit 11, acquisition unit 12, learning unit 13, and integration unit 14.

[0021] Alternatively, the communication establishment unit 11, acquisition unit 12, learning unit 13, and integration unit 14 may each be implemented with dedicated hardware. Furthermore, some or all of the components of each device may be implemented by general-purpose or dedicated circuits, processors, etc., or combinations thereof. These may be configured by a single chip or by multiple chips connected via a bus. Some or all of the components of each device may be implemented by a combination of the aforementioned circuits, etc., and programs. Furthermore, a CPU (Central Processing Unit), GPU (Graphics Processing Unit), FPGA (field-programmable gate array), etc., can be used as the processor.

[0022] Furthermore, if some or all of the components of the learning device 1 are realized by multiple information processing devices or circuits, these multiple information processing devices or circuits may be centrally located or distributed. For example, the information processing devices or circuits may be realized in a form in which each is connected via a communication network, such as a client-server system or a cloud computing system. Also, the functions of the learning device 1 may be provided in SaaS (Software as a Service) format.

[0023] The learning device according to Embodiment 1 establishes secure communication with information terminals connected to the networks of each organization and acquires datasets using this secure communication. Therefore, according to Embodiment 1, a global model can be constructed even when the networks of multiple organizations are not constantly connected.

[0024] <Embodiment 2> Embodiment 2 is a specific example of Embodiment 1. Figure 2 is a schematic diagram showing the configuration of the learning system 100 according to Embodiment 2. The learning system 100 comprises information terminals 2a, 2b, and 2c, VPN devices 3a, 3b, and 3c, and a learning device 4. The learning device 4 is a specific example of the learning device 1 described above.

[0025] Organization A's network Na has information terminal 2a and VPN device 3a. Organization B's network Nb has information terminal 2b and VPN device 3b. Organization C's network Nc has information terminal 2c and VPN device 3c.

[0026] Datasets owned by organization A are stored on information terminal 2a. Datasets owned by organization B are stored on information terminal 2b. Datasets owned by organization C are stored on information terminal 2c.

[0027] The number of organizations is not limited to three. There may be two organizations, or four or more. Each organization could be, for example, a pharmaceutical manufacturer or a chemical manufacturer. In this case, the dataset is a dataset of compounds. Each record in the compound dataset contains information such as the compound's structure and properties. The compound's structure is represented by a fixed-length bit string, where each bit in the bit string represents the presence or absence of a specific structure (e.g., a benzene ring). The property values ​​(e.g., tensile strength) may be values ​​obtained experimentally, or values ​​obtained through simulation or theoretical calculation. For example, data generated daily in the research and development work of organization A is stored in information terminal 2a. Of course, the dataset is not limited to a dataset of compounds; it can be a dataset of any object.

[0028] When information terminals 2a, 2b, and 2c are not distinguished from each other, they may simply be referred to as information terminal 2. When networks Na, Nb, and Nc are not distinguished from each other, they may simply be referred to as network N. Network N may be a LAN (Local Area Network), or a network connecting multiple LANs. Network N is connected to a public network PN, such as the Internet.

[0029] VPN device 3 is a VPN server or a VPN-compatible router. VPN device 3 may have the IP (Internet Protocol) address of learning device 4 pre-configured. The VPN may be an Internet VPN, an IP-VPN, or a wide-area Ethernet. When VPN devices 3a, 3b, and 3c are not distinguished from each other, they may simply be referred to as VPN device 3.

[0030] Figure 3 is a block diagram illustrating the configuration of the learning device 4. The learning device 4 is connected to the network PN. The learning device 4 comprises a communication establishment unit 41, an acquisition unit 42, a learning unit 43, and an integration unit 44.

[0031] The learning device 4 includes storage for storing local models La, Lb, and Lc. Local model La is a local model trained on a dataset owned by organization A. Local model Lb is a local model trained on a dataset owned by organization B. Local model Lc is a local model trained on a dataset owned by organization C. Local models La, Lb, and Lc are repeatedly updated by the learning unit 43. When local models La, Lb, and Lc are not distinguished from each other, they may simply be referred to as local model L.

[0032] The communication establishment unit 41 is a specific example of the communication establishment unit 11 described above. The communication establishment unit 41 establishes secure communication with the information terminal 2. Specifically, the communication establishment unit 41 connects to a VPN device 3, such as a VPN server, via the public network PN, and requests a VPN connection from the VPN device 3. First, a TCP / IP connection is established between the learning device 4 and the VPN device 3. Then, the learning device 4 is authenticated, and a VPN session is established between the learning device 4 and the VPN device 3. After the acquisition unit 42 acquires the dataset, the communication establishment unit 41 terminates the VPN session. The learning device 4 may also be connected to network N via a remote access VPN.

[0033] The timing at which the communication establishment unit 41 establishes secure communication, that is, the timing at which the learning device 4 establishes a VPN connection to network N, will be described later. This is because the timing may be related to the progress of processing in the learning unit 43, which will be described later. The timing at which secure communication is established with information terminal 2a, with information terminal 2b, and with information terminal 2c may be different from each other.

[0034] The acquisition unit 42 is a specific example of the acquisition unit 12 described above. The acquisition unit 42 acquires a dataset from the information terminal 2 after the learning device 4 has established a VPN connection to the network N.

[0035] The learning unit 43 is a specific example of the learning unit 13 described above. The learning unit 43 trains the corresponding local model L with the dataset acquired by the acquisition unit 42.

[0036] The integration unit 44 is a specific example of the integration unit 14 described above. The integration unit 44 integrates the local models La, Lb, and Lc that have been trained by the learning unit 43. The integrated model is called the global model. The integration unit 44 may integrate the local models La, Lb, and Lc at predetermined intervals (e.g., once a day, once every few months). The global model is more powerful than the local models La, Lb, and Lc. The integration unit 44 may also perform the integration of the local models La, Lb, and Lc after the training of the local models La, Lb, and Lc is complete.

[0037] The integration unit 44 may generate a global model by, for example, taking the arithmetic mean of the model parameters of local model La, local model Lb, and local model Lc. Note that the method of integrating the model parameters is not limited to the arithmetic mean.

[0038] After the integration unit 44 generates the global model, the learning device 4 distributes the global model to the information terminals 2a, 2b, and 2c. For example, after the process of generating the global model is completed, the learning device 4 may sequentially establish VPN connections to networks Na, Nb, and Nc and transmit the global model to the information terminals 2a, 2b, and 2c.

[0039] Furthermore, the learning device 4 may connect to network N via VPN in response to requests from each information terminal 2 and transmit the global model to the information terminal 2. Each information terminal 2 can import the global model at any time. Organizations A, B, and C will be able to utilize a high-performance global model that links datasets owned by multiple organizations.

[0040] The process of constructing multiple local models L and integrating them is also known as federative learning. In this case, it can be said that the learning device 4 is performing federative learning. However, it should be noted that the construction of local models L on local terminals such as information terminals 2a, 2b, and 2c is also sometimes referred to as federative learning. In Embodiment 2, the learning device 4 constructs the local models L.

[0041] The learning device 4 sequentially repeats the processes of establishing secure communication, acquiring a dataset, and training a local model with the acquired dataset. This allows the performance of the global model to be improved based on the datasets accumulated daily on each information terminal 2. The process of integrating multiple local models may be performed at any time.

[0042] Next, the timing of when the communication establishment unit 41 establishes secure communication will be explained. The communication establishment unit 41 may establish secure communication at a predetermined timing. This predetermined timing may be once every few months or once every few days.

[0043] Furthermore, the communication establishment unit 41 may establish secure communication in response to the receipt of requests from each information terminal 2. For example, the information terminal 2 sends a request when the amount of accumulated data sets exceeds a predetermined amount.

[0044] The communication establishment unit 41 may establish the next secure communication based on the progress of training the local model L with the dataset. When training the local model L with a single dataset, the dataset is divided into multiple batches, and the local model L is trained on each batch sequentially. The process of dividing the dataset into batches and training on each batch is repeated a predetermined number of times. The predetermined number of times is set so that the model parameters of the local model L converge. However, the predetermined number of times must be set to a number that does not cause overfitting. When the model parameters of the local model have converged, the communication establishment unit 41 may establish the next secure communication.

[0045] The progress of learning may be expressed as the number of learning iterations or the number of learned batches. For example, if the dataset is divided into 5 batches and learning is repeated 10 times, the next secure communication may be established when learning is complete, that is, when the 10th learning is finished. The communication establishment unit 41 may establish the next secure communication when learning is nearing completion, for example, when the 4th batch of the 10th learning is completed.

[0046] The communication establishment unit 41 may sequentially establish secure communication with information terminals 2a, 2b, and 2c when the learning progress of local model La, local model Lb, and local model Lc exceeds a threshold. Furthermore, the communication establishment unit 41 may establish secure communication with the corresponding information terminal 2 when the learning progress of any of the local models L exceeds a threshold.

[0047] The communication establishment unit 41 may establish the next secure communication based on the progress of the process of integrating multiple local models L. If the processing in the integration unit 44 is not a simple arithmetic mean, or if there are many organizations, the processing in the integration unit 44 may take a long time. It is efficient to start the next process after the processing in the integration unit 44 is completed.

[0048] Furthermore, applying secure computation techniques may require a long processing time for the integration unit 44. It is known that the dataset used for training can be inferred by reverse engineering the local model L. Therefore, to improve the confidentiality of the local model L, it is desirable to perform secure computation when integrating the local model L. Secure computation is a technique that performs computational processing while the data remains encrypted, and secure computation techniques using multi-party computation (MPC) and homomorphic encryption are known.

[0049] Figure 4 is a flowchart showing the process flow for generating the local model L. Assume that the learning device 4 has stored the initial local model L (step S101).

[0050] Next, the communication establishment unit 41 of the learning device 4 determines whether it is time to establish secure communication (step S102). If it is not time to establish secure communication (NO in step S102), the process returns to step S102.

[0051] If it is time to establish secure communication (YES in step S102), the communication establishment unit 41 establishes secure communication between the information terminal 2 and the learning device 4, and the acquisition unit 42 acquires the data set from the information terminal 2 (step S103). After that, the communication establishment unit 41 terminates the secure communication.

[0052] In step S103, multiple datasets may be acquired. First, secure communication is established between the information terminal 2a and the learning device 4, the acquisition unit 42 acquires a dataset from the information terminal 2a, and the communication establishment unit 41 terminates the secure communication. Then, secure communication is established between the information terminal 2b and the learning device 4, the acquisition unit 42 acquires a dataset from the information terminal 2b, and the communication establishment unit 41 terminates the secure communication. Then, secure communication is established between the information terminal 2c and the learning device 4, the acquisition unit 42 acquires a dataset from the information terminal 2c, and the communication establishment unit 41 terminates the secure communication. Of course, in step S103, a dataset may be acquired from any of the information terminals 2a, 2b, and 2c.

[0053] Next, the learning unit 43 trains the local model L with the dataset acquired in step S103 and updates the local model L (step S104). If multiple datasets were acquired in step S103, multiple local model L may be updated in step S104. After updating the local model L, the process returns to step S102. Note that the process of integrating multiple local model L may be performed at any time.

[0054] The learning device according to Embodiment 2 connects to each organization's network via VPN at an appropriate communication timing and acquires the organization's dataset. This allows for the secure reception of the dataset and the construction of a local model at the appropriate time.

[0055] Secure communication is not limited to VPN communication. Secure communication may be any secure communication protocol (e.g., encryption protocol). The dataset may also be sent from information terminal 2 to learning device 4 by email using a secure communication protocol (e.g., S / MIME).

[0056] <Modified form of Embodiment 2> The device equipped with the integration unit 44 that integrates the global model and the device equipped with the learning unit 43 that constructs the local model L may be different. In this case, the device equipped with the integration unit 44 may establish secure communication (e.g., SSL) with the device equipped with the learning unit 43 and obtain the local model L. This not only secures communication between the repository where the dataset is stored (e.g., information terminal 2) and the local model L, but also secures communication between the local model L and the global model.

[0057] <Embodiment 3> Embodiment 3 is a specific example of Embodiment 2. The learning device according to Embodiment 3 integrates the model parameters of the local model using secure computation. Figure 5 is a block diagram showing the configuration of the learning system 100a according to Embodiment 3. Comparing Figure 2 and Figure 5, a group of servers 5 has been added.

[0058] Server group 5 includes multiple secure computation servers 51. The number of secure computation servers 51 is not limited to three. However, it is preferable to have three or more secure computation servers 51 for efficient secure computation.

[0059] Server group 5 integrates local model La, local model Lb, and local model Lc, and transmits the result of the secure computation to learning device 4.

[0060] The integration unit 44 of the learning device 4 divides the model parameters of local model La into multiple (e.g., 3) shares and sends the multiple shares to multiple secure computing servers 51. The integration unit 44 divides the model parameters of local model Lb into multiple shares and sends the multiple shares to multiple secure computing servers 51. The integration unit 44 divides the model parameters of local model Lc into multiple shares and sends the multiple shares to multiple secure computing servers 51.

[0061] Each secure computation server 51 uses the received shares to perform secure computation to compute the global model. The shares do not reveal the local model, and computations using the shares can be considered secure computations. Multiple secure computation servers 51 may cooperate to perform multi-party computation (MPC). Since the computational complexity required to integrate the local model L is sufficiently small, the server group 5 is expected to be able to perform secure computations in a realistic amount of time.

[0062] Embodiment 3 also achieves the same effects as Embodiment 2. Furthermore, Embodiment 3 allows for the confidentiality of calculations for integrating the global model.

[0063] The program described above includes, when loaded into a computer, a set of instructions (or software code) for causing the computer to perform one or more of the functions described in the embodiments. The program may be stored in a non-temporary computer-readable medium or a physical storage medium. Examples, but not limited to, include random-access memory (RAM), read-only memory (ROM), flash memory, solid-state drive (SSD) or other memory technologies, CD-ROM, digital versatile disc (DVD), Blu-ray® disc or other optical disc storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage devices. The program may be transmitted over a temporary computer-readable medium or a communication medium. Examples, but not limited to, include temporary computer-readable medium or a communication medium that includes electrically, optically, acoustically, or otherwise propagating signals.

[0064] Although the present invention has been described above with reference to embodiments, the present invention is not limited thereto. Various modifications to the structure and details of the present invention can be made that are understandable to those skilled in the art within the scope of the invention.

[0065] Some or all of the above embodiments may also be described as follows, but are not limited to the following: (Note 1) A means for establishing secure communication with information terminals located on the network of each organization, An acquisition means that uses the aforementioned secure communication to acquire a dataset for each organization from the information terminal, A learning method for training a local model using the aforementioned dataset, An integration method for combining multiple local models that have been trained on multiple datasets. A learning device equipped with the following features. (Note 2) The communication establishment means establishes the next secure communication based on the learning progress of the local model. The learning device described in Appendix 1. (Note 3) The communication establishment means establishes the next secure communication when the model parameters of the local model converge. The learning device described in Appendix 1. (Note 4) The communication establishment means establishes the secure communication at a predetermined timing. The learning device described in Appendix 1. (Note 5) The communication establishment means establishes the secure communication in response to the receipt of a request from each information terminal. The learning device described in Appendix 1. (Note 6) When the amount of data in the dataset stored in the information terminal exceeds a predetermined amount, the information terminal transmits the request. The learning device described in Appendix 5. (Note 7) The communication establishment means establishes the next secure communication based on the progress of the process of integrating the multiple local models. The learning device described in Appendix 1. (Note 8) The integration means integrates the multiple local models using secure computation technology. The learning device described in Appendix 7. (Note 9) The communication establishment means establishes the secure communication by connecting the learning device to the network via VPN (Virtual Private Network). A learning device as described in any one of the items 1 to 8 of the appendix. (Note 10) Information terminals located on each organization's network, Learning device, A learning system equipped with, The learning device is Secure communication is established with the aforementioned information terminal, Using the aforementioned secure communication, obtain organization-specific datasets from the information terminal. The aforementioned dataset is used to train a local model. Integrate multiple local models trained on multiple datasets. Learning system. (Note 11) The learning device establishes the following secure communication based on the learning progress in the local model. The learning system described in Appendix 10. (Note 12) Computers Secure communication is established between information terminals located on each organization's network. Using the aforementioned secure communication, obtain organization-specific datasets from the information terminal. The aforementioned dataset is used to train a local model. Integrate multiple local models trained on multiple datasets. Learning methods. (Note 13) On the computer, The process of establishing secure communication with information terminals located on each organization's network, The process of obtaining organization-specific datasets from information terminals using the aforementioned secure communication, The process of training a local model using the aforementioned dataset, The process of integrating multiple local models that have been trained on multiple datasets and A non-temporary, computer-readable medium that stores a program to execute. [Explanation of Symbols]

[0066] 1.4 Learning device 11, 41 Communication Establishment Unit 12, 42 Acquisition Department 13, 43 Learning Department 14, 44 Integration Department 2, 2a, 2b, 2c Information terminals 3, 3a, 3b, 3c VPN device 100, 100a Learning System 5 Server Clusters 51 Secure Computing Server N, Na, Nb, Nc Network PN Public Network

Claims

1. A means for establishing secure communication with information terminals located on the network of each organization, An acquisition means that uses the aforementioned secure communication to acquire a dataset for each organization from the information terminal, A learning method for training a local model using the aforementioned dataset, An integration method for combining multiple local models that have been trained on multiple datasets. A learning device equipped with the following features.

2. The communication establishment means establishes the next secure communication after the end of the previous secure communication, based on the progress of the learning of the local model. The learning device according to claim 1.

3. The communication establishment means establishes the next secure communication after the end of the previous secure communication when the model parameters of the local model have converged. The learning device according to claim 1.

4. The communication establishment means establishes the secure communication at a predetermined timing. The learning device according to claim 1.

5. The communication establishment means establishes the secure communication in response to the receipt of a request from each information terminal. The learning device according to claim 1.

6. The aforementioned request is transmitted when the amount of data in the dataset stored in the information terminal exceeds a predetermined amount. The learning device according to claim 5.

7. The communication establishment means establishes the next secure communication after the completion of the previous secure communication, based on the progress of the process of integrating the multiple local models. The learning device according to claim 1.

8. The integration means integrates the multiple local models using secure computation technology. The learning device according to claim 7.

9. The communication establishment means establishes the secure communication by connecting the learning device to the network via VPN (Virtual Private Network). A learning device according to any one of claims 1 to 8.

10. Information terminals located on each organization's network, Learning device, A learning system equipped with, The learning device is Secure communication is established with the aforementioned information terminal, Using the aforementioned secure communication, obtain organization-specific datasets from the information terminal. The aforementioned dataset is used to train a local model. Integrate multiple local models trained on multiple datasets. Learning system.

11. The learning device establishes the next secure communication after the end of the previous secure communication, based on the progress of learning in the local model. The learning system according to claim 10.

12. Computers Secure communication is established between information terminals located on each organization's network. Using the aforementioned secure communication, obtain organization-specific datasets from the information terminal. The aforementioned dataset is used to train a local model. Integrate multiple local models trained on multiple datasets. Learning methods.

13. On the computer, The process of establishing secure communication with information terminals located on each organization's network, The process of obtaining organization-specific datasets from information terminals using the aforementioned secure communication, The process of training a local model using the aforementioned dataset, The process of integrating multiple local models that have been trained on multiple datasets and A program to execute.

Images