Learning system, learning device, and learning program

The learning system quantifies client data uniqueness to enhance federated learning accuracy by optimizing model influence, addressing privacy concerns and improving prediction performance.

JP2024161786AActive Publication Date: 2024-11-20DENSO CORP +1
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
JP2023076809
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-08
Publication Date
2024-11-20
Estimated Expiration
2043-05-08

AI Technical Summary

Technical Problem

Federated learning systems face challenges in achieving high estimation accuracy while protecting privacy, as sharing a deep learning network model leads to high communication loads and complicates data understanding due to the exchange of large data sizes.

Method used

A learning system that quantifies the uniqueness of client data using explainability and predictive divergence, allowing clients to learn individual models based on server parameters without sharing actual data, optimizing the influence of the general model according to uniqueness, and updating both server and client models.

Benefits of technology

Enables high estimation accuracy in federated learning without compromising privacy, by understanding data characteristics and improving prediction accuracy through optimized model influence and unique data recognition.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024161786000001_ABST
    Figure 2024161786000001_ABST
Patent Text Reader

Abstract

To provide a learning system capable of achieving high estimation accuracy without impairing characteristics such as protection of privacy in association learning.SOLUTION: A learning system 1 including a server 10 and a plurality of clients 20 each connected to the server 10 so as to communicate with each other has a joint learning function using data 23a, 23b held by each of the plurality of clients 20. The client 20 includes: a general model 21 configured by using a parameter of a server model 11 owned by the server 10; an individual model 22 configured to be capable of learning by using the general model 21 and the data 23a, 23b held by the client 20; an originality quantifying unit 24 for quantifying originality of the data 23a, 23b owned by the client 20 relative to the general model 21; and a model learning unit 25 for learning the individual model 22 by changing influence of the general model according to originality.SELECTED DRAWING: Figure 1
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The disclosure of this specification relates to a technique for performing federated learning using unique data each owned by multiple clients. [Background technology]

[0002] In the learning system disclosed in Patent Document 1, in order to protect privacy in federated learning, a part of the deep learning network of a specific model owned by a client is shared as a common model by multiple clients and a server. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2022-78973 Summary of the Invention [Problem to be solved by the invention]

[0004] However, as in Patent Document 1, when a part of a deep learning network is shared as a common model, a special structure and management are required to share the common model part. In addition, since the model itself, which has a large data size, is shared, a high communication load occurs. In this way, the learning system of Patent Document 1 may cause complication in federated learning. In federated learning, it is required to achieve high estimation accuracy without compromising characteristics such as privacy protection, and at the same time, since only parameters are exchanged between the server and the client and data is not exchanged directly, it is difficult to grasp the characteristics of the data of each client, and there is also a problem that accuracy cannot be improved by using data characteristics.

[0005] One of the objectives of the disclosure of this specification is to provide a learning system, learning device, and learning program that can extract the characteristics of the data held by each client in federated learning without compromising characteristics such as privacy protection, and that can achieve high estimation accuracy using the extracted distribution characteristics. [Means for solving the problem]

[0006] One aspect disclosed herein is a learning system having a server (10) and a plurality of clients (20) each communicatively connected to the server, the learning system having a federated learning function using data (23a, 23b) held by each of the plurality of clients, The server has a server model (11) configured to be able to learn based on information collected from each client, The client: A general model (21) constructed using parameters of the server model; An individual model (22) configured to be trainable using the general model and data held by the client; a uniqueness quantification unit (24) for quantifying the uniqueness of the data held by the client with respect to the general model; and a model learning unit (25) for learning an individual model by changing the influence of the general model in accordance with the quantified uniqueness.

[0007] Another disclosed aspect is a learning device for configuring one of the clients (20a) in a learning system (1) having a server (10) and a plurality of clients (20) each connected to the server so as to be able to communicate with the server, the learning device having a federated learning function using data held by each of the plurality of clients, the learning device comprising: A general model (21) that is constructed by acquiring parameters of a server model (11) provided in a server and using the parameters of the server model; An individual model (22) configured to be able to learn using parameters of the server model and data (23a, 23b) held by the client itself; A uniqueness quantification unit (24) that quantifies the uniqueness of data held by the client with respect to the general model; The system further includes a model learning unit (25) for learning the individual model by changing the influence of the general model in accordance with the quantified uniqueness.

[0008] Another disclosed aspect is a learning program configured in a client (20a) communicably connected to a server (10) for performing federated learning with another client (20b) via the server, the learning program comprising: At least one processor has Acquiring parameters of a server model (11) provided in a server, and generating a general model (21) configured using the parameters of the server model and an individual model (22) configured to be able to learn using the parameters of the server model and data (23a, 23b) held by the client itself; Quantifying the uniqueness of the data held by our client relative to the general model; and The method is configured to change the influence of the general model according to the quantified uniqueness to train the individual model; and

[0009] According to these aspects, the uniqueness of the data held by the client is quantified in the client, so that the characteristics of the data held by the client can be understood without directly sharing and analyzing each piece of data that requires privacy consideration. Therefore, in federated learning, high estimation accuracy can be achieved without compromising characteristics such as privacy protection.

[0010] In addition, the symbols in parentheses included in the claims, etc., are intended to exemplify the correspondence with parts of the embodiments described below, and are not intended to limit the technical scope. [Brief description of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram showing a schematic configuration of a learning system. [Diagram 2] FIG. 11 is a diagram showing an example of processing by the learning system. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] An embodiment will be described with reference to the drawings.

[0013] (First embodiment) As shown in FIG. 1, the learning system 1 according to the first embodiment includes a server 10 and a plurality of clients 20. The learning system 1 has a federated learning function using data 23a, 23b held by each of the plurality of clients 20. In federated learning, for example, the server 10 may collect the data 23a, 23b held by each of the clients 20 and execute learning on the server 10 side. However, the data 23a, 23b held by the client 20 may include personal information, customer information, confidential information, etc., and there may be cases where there is a right to privacy. For this reason, in the learning system 1, the data 23a, 23b itself is not communicated between the server 10 and each of the clients 20, but parameters defining the learning model are the subject of communication.

[0014] The learning model is a model that has parameters that can be learned based on predetermined data. For example, a neural network may be adopted as the learning model. The neural network may be a convolutional neural network or various other neural networks. For example, when a convolutional neural network is adopted as the learning model, the parameters that define the learning model may be parameters set in a local filter applied to a convolution layer. For the learning model, a model other than a neural network, for example, a recurrent neural network model, a tree-structured / ensemble learning model such as Random Forest, a regression model, etc. may be adopted. The learning models adopted in the learning system 1 may be models having a common structure so that compatibility occurs in the application of parameters.

[0015] The server 10 is mainly composed of a computer and has at least one memory and one processor. The memory may be at least one type of non-transient substantial storage medium, such as a semiconductor memory, a magnetic medium, or an optical medium, which non-temporarily stores programs and data that can be read by the processor. Further, a rewritable volatile storage medium, such as a random access memory (RAM), may be provided as the memory. The processor includes at least one type of core, such as a central processing unit (CPU), a graphics processing unit (GPU), or a reduced instruction set computer (RISC)-CPU.

[0016] The server 10 has a plurality of functional units for implementing an associative learning function by executing a program by a processor. Specifically, the server 10 includes a server model 11, a communication unit 12, and a model update unit 13.

[0017] The server model 11 is a learning model that is optimized for common data 23 a that commonly appears among the clients 20 .

[0018] The communication unit 12 is configured to be able to transmit parameters of the server model 11 to each client 20 through the communication network. The communication unit 12 is also configured to be able to receive parameters necessary for updating the server model 11 from each client 20 through a network line. Furthermore, the communication unit 12 may be configured to be able to receive information regarding the uniqueness of data (described later) in association with parameters necessary for updating the server model 11. The communication network may be a wired network or a wireless network.

[0019] The model update unit 13 is configured to be able to update the server model 11 to the latest state by using parameters necessary for updating the server model 11 received from each client 20. When updating the server model 11, the model update unit 13 may take into consideration the uniqueness of the data 23a, 23b held by each client 20.

[0020] The multiple clients 20 have common functions and structures to the extent described below, so the following description will be directed to the client 20a as a representative of the multiple clients 20.

[0021] The client 20a is a system including a learning device mainly composed of a computer. The configuration of the client 20a described below can be interpreted as the configuration of a learning device provided in the client 20a. The client 20a has at least one memory and one processor. The memory may be at least one type of non-transient substantial storage medium, such as a semiconductor memory, a magnetic medium, or an optical medium, which non-temporarily stores programs and data that can be read by the processor. Furthermore, a rewritable volatile storage medium, such as a random access memory (RAM), may be provided as the memory. The processor includes at least one type of core, such as a central processing unit (CPU), a graphics processing unit (GPU), or a reduced instruction set computer (RISC)-CPU.

[0022] The client 20a has a plurality of functional units for implementing an associative learning function by executing a program (e.g., a learning program) by a processor. Specifically, the client 20a includes a general model 21, an individual model 22, a uniqueness quantification unit 24, a model learning unit 25, a communication unit 26, and a display unit 27.

[0023] In addition, the client 20a holds data 23a and 23b that can be teacher data used in the associative learning. The data 23a and 23b held by the client 20a can be classified into common data 23a that appears in common among multiple clients 20 and unique data 23b that appears uniquely in a specific client. However, in the present learning system 1, it is not necessary to share the data itself between the server 10 and multiple clients 20. For this reason, it is sufficient that the overall tendency of the data 23a and 23b is analyzed by quantifying the uniqueness of the individual model 22 described later, and the learning system 1 does not need to individually distinguish the actual data into the common data 23a and the unique data 23b.

[0024] The general model 21 is provided to update the server model 11. The general model 21 is a model configured using server-side parameters acquired from the server model 11.

[0025] The individual model 22 is a model provided uniquely to the client 20a. The individual model 22 is configured to be able to learn using the general model 21 and data 23a and 23b held by the client 20a. Since the individual model 22 is learned individually by the client 20a, the parameters of the individual model 22 are different from the parameters of the server model 11 and the parameters of the general model 21.

[0026] The uniqueness quantification unit 24 quantifies the uniqueness of the data 23a, 23b held by the client 20a. As indicators representing the uniqueness of the data 23a, 23b, the explainability by the general model 21 (hereinafter simply referred to as explainability) and the prediction deviation between the individual model 22 and the general model 21 (hereinafter simply referred to as prediction deviation) are adopted. The uniqueness quantification unit 24 calculates the explainability and the prediction deviation.

[0027] The explainability is an index indicating whether the data 23a, 23b can be discriminated (explained) by the general model 21. As an example, the explainability is expressed by the following formula 1 using cross-entropy error.

[0028]

number

[0029]

number

[0030]

number

[0031]

number

[0032] The prediction divergence is an index showing whether the predicted values ​​are the same when the data 23a, 23b held by the client 20a are evaluated by the individual model 22 and the general model 21, respectively. In order to quantify the index, the uniqueness quantification unit 24 calculates the difference between the predicted value of the individual model 22 and the predicted value of the general model 21 by, for example, the formula shown in the following formula 5 using the KL divergence. The KL divergence is a measure for measuring the difference in probability distribution.

[0033]

number

[0034]

number

[0035]

number

[0036]

number

[0037]

number

[0038]

number

[0039]

number

[0040]

number

[0041] Here, the model learning unit 25 defines classification loss and compatibility loss as two loss functions. The loss function of classification loss is expressed by the following formula 13 using cross-entropy error. The classification loss is calculated by dividing each of the data 23a and 23b by uniqueness m i By multiplying by , the value becomes larger when the common data 23a is learned and becomes smaller when the unique data 23b is learned.

[0042]

number

[0043]

number

[0044]

number

[0045] Furthermore, the model learning unit 25 may have a function of learning the general model 21. The loss function adopted in learning the general model 21 may be expressed by the following mathematical expression 16 using cross-entropy error.

[0046]

number

[0047]

number

[0048]

number

[0049] The display unit 27 has a function of displaying information regarding the uniqueness of data held by the client 20a. This function may be a function of displaying on a display device such as a liquid crystal display device provided in the client 20a, or a function of transmitting information regarding the uniqueness to a terminal such as a smartphone communicably connected to the client 20a and displaying the information on a display screen of the terminal.

[0050] The information on uniqueness may include the quantified numerical value of the uniqueness calculated by the uniqueness quantification unit 24. The information on uniqueness may include a graph that visualizes the quantified numerical value. The information on uniqueness may include a sentence that is automatically generated using artificial intelligence or the like based on the quantified numerical value.

[0051] In addition, when the client 20a is regarded as the own client, the client 20b, which is regarded as another client, has the same functions and structure as the client 20a. However, there are differences between the client 20a and the data 23a and 23b held by the client 20b and the parameters of the individual model 22 updated using the data.

[0052] Next, an example of a method of associative learning by the learning system 1 will be described with reference to the flowchart in Fig. 2. Note that the method of associative learning using uniqueness is not limited to this method, and various methods can be adopted within the scope of the present disclosure.

[0053] 2, in S1, the server 10 transmits parameters of the server model 11 to each client 20. After the processing of S1, the process proceeds to S2.

[0054] In S2, each client 20 generates a general model 21 using the parameters of the server model 11 received by each client 20. After the process of S2, the process proceeds to S3.

[0055] In S3, each client 20 uses the data 23a, 23b that it holds to generate its own individual model 22. After the process of S3, the process proceeds to S4.

[0056] In S4, each client 20 (particularly the uniqueness quantification unit 24) quantifies the uniqueness of the data 23a, 23b held by each client 20 using the general model 21 and the individual model 22. After the process of S4, the process proceeds to S5.

[0057] In S5, each client 20 (particularly, the model learning unit 25) executes learning using the general model 21 and the quantified uniqueness, and updates the respective individual models 22. After the process of S5, the process proceeds to S6.

[0058] In S6, each client 20 transmits to the server 10 parameters and the like required for updating the server model 11. After the processing of S6, the process proceeds to S7.

[0059] In S7, the server 10 updates the server model 11 using the data received from each client 20. The series of processes ends with S7.

[0060] The processes of S1 to S6 may be executed at different times for each client.

[0061] According to the first embodiment described above, the uniqueness of the data 23a, 23b held by the client 20a is quantified, so that the client 20a can understand the characteristics of the data 23a, 23b held by the client 20a without directly sharing the data 23a, 23b whose privacy should be considered. Therefore, in the federated learning, it is possible to achieve high estimation accuracy without compromising characteristics such as privacy protection.

[0062] Moreover, according to the first embodiment, in learning of the individual model 22 using the general model 21 and the data 23a, 23b held by the client 20a, the influence of the general model 21 in learning is changed according to uniqueness. Since the influence of the general model 21 can be optimized according to the characteristics of the individual model 22 and reflected in the individual model 22, the prediction accuracy of the individual model 22 on the client 20a side or the accuracy of processing using the individual model 22 can be improved.

[0063] Moreover, according to the first embodiment, uniqueness is quantified using the explainability of the data 23a, 23b held by the client 20a by the general model 21 and the prediction dissociability between the individual model 22 and the general model 21. Since uniqueness is evaluated using two types of indices, the accuracy of associative learning can be further improved.

[0064] Moreover, according to the first embodiment, the explainability of the data 23a, 23b held by the client 20a by the general model 21 is calculated based on the prediction accuracy of the server model 11, since the general model 21 is configured using parameters of the server model 11. That is, the uniqueness of the data 23a, 23b held by the client 20a with respect to the common data 23a common to a plurality of clients 20 can be appropriately estimated.

[0065] Moreover, according to the first embodiment, the prediction dissociation between the individual model 22 and the general model 21 is calculated based on the difference between the prediction results of the individual model 22 and the general model 21. That is, it is possible to appropriately estimate the uniqueness of the data 23a, 23b held by the client 20a.

[0066] In this way, in the first embodiment, it is possible to quantify whether the data 23a, 23b held by each client is unique data 23b or common data 23a. Furthermore, by using an index of uniqueness, it is expected to improve the learning efficiency (in other words, the convergence speed) and prediction performance of both the individual model 22 on the client 20 side and the server model 11 on the server 10 side.

[0067] (Other embodiments) Although one embodiment has been described above, the present disclosure should not be construed as being limited to this embodiment, and can be applied to various embodiments within the scope not departing from the gist of the present disclosure.

[0068] For example, a display unit that displays information about uniqueness may be provided in the server 10. In this case, the display unit may be configured to be capable of displaying the uniqueness of each client in a list.

[0069] Also, for example, it is preferable that all clients 20 that are the subject of federated learning and connected to the server 10 are provided with a function for quantifying uniqueness, but exceptionally, some clients may not be provided with a function for quantifying uniqueness.

[0070] For example, federated learning using uniqueness can be used in in-vehicle network solutions in which the client 20 is a vehicle or an in-vehicle device, or in medical network solutions in which the client 20 is a terminal installed in a medical institution.

[0071] The controller and method described herein may be implemented by a special-purpose computer having a processor programmed to execute one or more functions embodied in a computer program. Alternatively, the device and method described herein may be implemented by a special-purpose hardware logic circuit. Alternatively, the device and method described herein may be implemented by one or more special-purpose computers configured by a combination of a processor executing a computer program and one or more hardware logic circuits. The computer program may also be stored on a computer-readable non-transitory tangible recording medium as instructions executed by the computer. [Explanation of symbols]

[0072] 1: Learning system, 10: Server, 11: Server model, 20, 20a, 20b: Client, 21: General model, 22: Individual model, 23a, 23b: Data, 24: Uniqueness quantification unit

Claims

1. A learning system comprising a server (10) and a plurality of clients (20) each communicably connected to the server, the learning system having an associative learning function using data (23a, 23b) held by each of the plurality of clients, The server has a server model (11) configured to be able to learn based on information collected from each of the clients; The client: A general model (21) constructed using parameters of the server model; An individual model (22) configured to be trainable using the general model and data held by the client; a uniqueness quantification unit (24) for quantifying the uniqueness of data held by the client with respect to the general model; A model learning unit (25) that changes the influence of the general model in accordance with the quantified uniqueness and learns the individual model.

2. The learning system according to claim 1 , wherein the uniqueness quantification unit quantifies the uniqueness using the explainability of the data held by the client by the general model and the predictive dissociation between the individual model and the general model.

3. The learning system of claim 2 , wherein the explainability is calculated based on a predictive accuracy of the server model.

4. The learning system according to claim 2 , wherein the prediction dissociation is calculated based on a difference between prediction results of the individual model and the general model.

5. 2. The learning system according to claim 1, wherein at least one of the server and each of the clients further comprises a display unit (27) for displaying information regarding the uniqueness.

6. The model learning unit further changes the influence of the individual model in accordance with the quantified uniqueness to learn the general model; The learning system according to claim 1 , wherein the server trains the server model based on parameters of the general model that has already been trained in each of the clients.

7. A learning device for configuring one of the clients (20a) in a learning system (1) having a federated learning function using data held by each of the clients, the learning device comprising: a server (10); and a plurality of clients (20) each connected to the server so as to be able to communicate with the server; the learning device configuring one of the clients (20a) in a learning system having a federated learning function using data held by each of the clients, the learning device comprising: A general model (21) that is constructed by acquiring parameters of a server model (11) provided in the server and using the parameters of the server model; An individual model (22) configured to be able to learn using parameters of the server model and data (23a, 23b) held by the client itself; a uniqueness quantification unit (24) that quantifies the uniqueness of the data held by the client with respect to the general model; A model learning unit (25) that changes the influence of the general model according to the quantified uniqueness and learns the individual model.

8. A learning program configured in a client (20a) communicably connected to a server (10) for performing federated learning with another client (20b) via the server, At least one processor, Acquiring parameters of a server model (11) provided in the server, and generating a general model (21) configured using the parameters of the server model and an individual model (22) configured to be able to learn using the parameters of the server model and data (23a, 23b) held by the client itself; Quantifying the uniqueness of the data held by the own client with respect to the general model; and training the individual model by varying the influence of the general model in response to the quantified uniqueness.

Citation Information

Patent Citations

  • Method and system for initializing a neural network

    JP2022531882A

  • Information processing apparatus, information processing system, information processing method, and program

    JP2023012757A

  • Navigation system

    JP2023514079A

  • Model learning system and model learning method

    JP2022078973A