Method and information processing device

Abstract

To reduce the amount of data flowing between information processing devices when training a machine learning model on multiple information processing devices.SOLUTION: A first information processing device transmits first information used to obtain parameter values of a machine learning model to multiple second information processing devices. The multiple second information processing devices obtain evaluation values of the machine learning model when the parameter values obtained on the basis of the first information are applied to the machine learning model and transmit the evaluation values to the first information processing device. The first information processing device aggregates the multiple evaluation values received from the multiple second information processing devices and transmits an aggregation result of the evaluation values to the multiple second information processing devices. The first information processing device and the multiple second information processing devices update the machine learning model on the basis of the aggregation result of the evaluation values.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present disclosure relates to the learning of machine learning models.

Background Art

[0002] Generally, when training a machine learning model with training data held by multiple devices, the training data is collected at one location, such as a server, and the machine learning model is trained on the server. However, in this case, the confidentiality of the training data is not protected. Federated learning is a method for training a machine learning model while protecting the confidentiality of the training data. Federated learning is a method of training a machine learning model while keeping the distributed training data distributed.

[0003] In federated learning, each client has a machine learning model, trains the machine learning model with the held training data, and sends the trained machine learning model to the server. The server merges the trained machine learning models collected from each client and sends the merged machine learning model back to the clients.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Non-Patent Documents

[0005]

Non-Patent Document 1

[0006] However, in federated learning, the machine learning model itself and / or the parameters of the machine learning model are transmitted between the server and each client, so the larger the size of the machine learning model, the greater the amount of data transmitted.

[0007] One aspect of the disclosure is to provide a method and an information processing device that can reduce the amount of data flowing between information processing devices when training a machine learning model using multiple information processing devices. [Means for solving the problem]

[0008] One aspect of this disclosure is, The first information processing device, The first information, used to obtain parameter values ​​for a machine learning model, is transmitted to multiple second information processing devices. The plurality of second information processing devices, The evaluation value of the machine learning model is obtained when the parameter values ​​obtained based on the first information are applied to the machine learning model. The evaluation value is transmitted to the first information processing device. The first information processing device, The multiple evaluation values ​​received from the multiple second information processing devices are aggregated, The aggregated results of the evaluation values ​​are transmitted to the plurality of second information processing devices. The first information processing device and the plurality of second information processing devices Based on the aggregated results of the evaluation values, update the machine learning model. It is a method.

[0009] Another aspect of this disclosure is, Transmitting first information used to obtain parameter values ​​for a machine learning model to multiple information processing devices, The plurality of information processing devices receive an evaluation value of the machine learning model when the parameter values ​​obtained based on the first information are applied to the machine learning model, The process involves aggregating the multiple evaluation values ​​received from the multiple information processing devices, The aggregated results of the evaluation values ​​are transmitted to the multiple information processing devices, The machine learning model is updated based on the aggregated results of the aforementioned evaluation values. A control unit that executes This is an information processing device equipped with [a specific feature / feature].

[0010] Another aspect of this disclosure is, The first information processing device receives first information used to obtain parameter values ​​for a machine learning model, Obtaining an evaluation value of the machine learning model when applying the value of the parameter obtained based on the first information to the machine learning model; Transmitting the evaluation value to the first information processing device; Receiving, from the first information processing device, an aggregation result of a plurality of the evaluation values by a plurality of information processing devices that have received the first information; Updating the machine learning model based on the aggregation result of the evaluation values; A control unit that executes; An information processing device comprising the same.

Advantages of the Invention

[0011] According to one aspect of the present disclosure, when learning a machine learning model with a plurality of information processing devices, the amount of data flowing between the information processing devices can be reduced.

Brief Description of the Drawings

[0012] [Figure 1] ]>It is a figure which shows an example of the system configuration of the federated learning system which concerns on 1st Embodiment. [Figure 2] It is a figure which shows an example of the structure of a server and a client. [Figure 3] It is a figure which shows an example of the sequence of processing in a federated learning system. [Figure 4] It is a figure which shows an example of the system configuration of the federated learning system which concerns on 2nd Embodiment.

Modes for Carrying Out the Invention

[0013] One aspect of the present disclosure is a method of learning a machine learning model with a first information processing device and a plurality of second information processing devices. The first information processing device is, for example, a server. The second information processing device is, for example, a PC (Personal Computer), a smartphone, a tablet terminal The first information processing device may be a terminal, for example, similar to the second information processing device. Examples of in-vehicle devices include data communication devices, car navigation systems, drive recorders, and other ECUs (Electronic Control Units).

[0014] The first information processing device transmits first information to multiple second information processing devices, which is used to obtain parameter values ​​for a machine learning model. The first information is a predetermined number of strings, such as a random number seed and a code associated with a parameter. The multiple second information processing devices obtain an evaluation value of the machine learning model when the parameter values ​​obtained based on the first information are applied to the machine learning model, and transmit this evaluation value to the first information processing device. For example, if the first information is a random number seed, the multiple second information processing devices generate random numbers from the received random number seed using a common random number generation method, and obtain the parameters of the machine learning model based on these random numbers. For example, if the first information is a code, the multiple second information processing devices obtain the parameter value corresponding to the received code from a pre-held correspondence table of codes and parameter values.

[0015] The first information processing device aggregates multiple evaluation values ​​of the machine learning model for the first information received from multiple second information processing devices, and transmits the aggregated results to the multiple second information processing devices. The first information processing device and the multiple second information processing devices update the machine learning model based on the aggregated evaluation values.

[0016] In one aspect of this disclosure, the data flowing between a first information processing device and a plurality of second information processing devices consists of first information and evaluation values ​​and aggregate values ​​of a machine learning model for the first information. The first information, evaluation values, and aggregate values ​​are all strings of a predetermined number. Therefore, the size of the first information, evaluation values, and aggregate values ​​is approximately a few bytes each. Accordingly, according to one aspect of this disclosure, when training a machine learning model using multiple information processing devices, the amount of communication data flowing between information processing devices can be reduced.

[0017] Furthermore, the machine learning model possessed by the first information processing device and the machine learning models possessed by each of the multiple second information processing devices may be the same model or different models. Also, the content of the first information transmitted to each of the multiple second information processing devices may be the same or different among the multiple information processing devices. In addition, each of the multiple second information processing devices may receive one piece of the first information, or multiple pieces of the first information.

[0018] Embodiments of this disclosure will be described below with reference to the drawings. The configurations of the following embodiments are illustrative, and this disclosure is not limited to the configurations of these embodiments.

[0019] <First Embodiment> Figure 1 shows an example of the system configuration of the federated learning system 100 according to the first embodiment. The federated learning system 100 is a system that trains a machine learning model when the training data is distributed across multiple devices. The federated learning system 100 includes a server 1 and multiple clients 2. The server 1 and the multiple clients 2 are each connected to a network N1 and can communicate with each other through the network N1. The network N1 is, for example, a public network such as the Internet, a LAN (Local Area Network), and a virtual or physical dedicated network.

[0020] In the first embodiment, it is assumed that server 1 and the multiple clients 2 have the same machine learning model. It is also assumed that server 1 and the multiple clients 2 share the parameter values ​​applied to the machine learning model. Therefore, When a new client 2 joins the federated learning system 100, Server 1 distributes the machine learning model and the latest parameter values ​​to the new client 2. Hereafter, when referred to as "parameters," it refers to the parameter values.

[0021] In the federated learning system 100, the following processes are performed: (1) Server 1 sends parameter generation information, which is information that serves as the seed for generating parameters, to each client 2. Parameter generation information includes, for example, a random number seed and a code associated with a parameter or a parameter variation value. Parameter generation information is a predetermined number of strings. Therefore, the size of the parameter generation information is, for example, about 1 to 100 bytes. Parameter generation information is an example of "first information".

[0022] (2) Each client 2 obtains parameters based on the received parameter generation information, obtains an evaluation value of the machine learning model when those parameters are applied, and sends that evaluation value to server 1.

[0023] The evaluation value represents the accuracy of a machine learning model when the parameters obtained based on the received parameter generation information are applied to the machine learning model. For example, in the case of supervised learning, the evaluation value may be a similarity metric indicating how close the output value of the machine learning model is to the correct answer. For example, in the case of reinforcement learning, the evaluation value may be a value indicating how well the output value of the machine learning model meets the required criteria. Hereinafter, the evaluation value of a machine learning model when the parameters obtained based on the parameter generation information are applied to the machine learning model will simply be referred to as the evaluation value for the parameter generation information. Since the evaluation value is a sequence of a predetermined number of numbers, its size is only a few bytes. The evaluation value of a machine learning model can also be said to be a value that represents the performance of the machine learning model. The performance of a machine learning model includes, for example, "accuracy," such as how close the output value of the machine learning model is to the correct answer, or how well the output value of the machine learning model meets the required criteria, and also includes the operational costs, such as human resources and expenses.

[0024] (3) Server 1 aggregates the evaluation values ​​of the parameter generation information received from each client 2 and sends the aggregated results to each client 2. (4) Server 1 and each client 2 update the machine learning models they each hold based on the aggregated evaluation values. In the first embodiment, updating a machine learning model means updating the parameter values ​​of the machine learning model. In the first embodiment, in (4), if the aggregated evaluation values ​​are better than the evaluation values ​​of the currently applied parameters, Server 1 and each client 2 update the machine learning model by applying the parameters obtained based on the parameter generation information sent from Server 1 in (1).

[0025] In the federated learning system 100, the machine learning model is trained by performing processes (1) to (4) for a predetermined number of random number seeds, or by repeatedly performing them until the aggregated evaluation values ​​satisfy a predetermined condition. Server 1 is an example of a "first information processing device". Client 2 is an example of a "second information processing device".

[0026] Each client 2 holds different evaluation data. Therefore, as in the first embodiment, even if each client 2 uses the same machine learning model and the same parameter values, the evaluation values ​​will differ for each client 2. However, by updating the parameter values ​​of the machine learning model with better parameter values ​​based on the aggregated evaluation values ​​of each client 2, the machine learning model becomes more accurate when viewed from the perspective of all clients 2. A machine learning model with an accuracy comparable to that of a normally trained machine learning model can be obtained.

[0027] In the federated learning system 100, the data flowing between server 1 and each client 2 is: This consists of parameter generation information, evaluation values, and aggregated evaluation values, each amounting to a few bytes. Therefore, according to the first embodiment, the amount of communication data can be reduced when training a machine learning model while the training data is distributed across multiple devices. In the following description, it is assumed that a random number seed is used as parameter generation information in the federated learning system 100.

[0028] Figure 2 shows an example of the configuration of Server 1 and Client 2. Server 1 is, for example, a computer dedicated to the federated learning system 100. However, Server 1 is not limited to this. Server 1's hardware configuration includes, for example, a processor 11, memory 12, auxiliary storage device 13, and a communication unit 14.

[0029] The auxiliary storage device 13 stores, for example, the OS (Operating System), several other programs, and data used by the processor 11 when executing each program. The auxiliary storage device 13 is, for example, an HDD (Hard Disk Drive) and an SSD (Solid State Drive). The memory 12 is a storage device that provides the processor 11 with a storage area and a working area for loading programs stored in the auxiliary storage device 13, and is also used as a buffer. The memory 12 includes, for example, semiconductor memory such as ROM (Read Only Memory) and RAM (Random Access Memory).

[0030] The processor 11 performs various processes by loading programs held in the auxiliary storage device 13 into the memory 12 and executing them. There is not limited to one processor 11; multiple processors may be provided. The processor 11 is an example of a "control unit".

[0031] The communication unit 14 is a module that connects, for example, a LAN (Local Area Network) card and network cables such as optical modules, and includes a signal processing circuit. The communication unit 14 is not limited to a circuit that can connect to a wired network, but may also be a wireless signal processing circuit that can process wireless signals from a wireless communication network such as WiFi.

[0032] The processor 11 executes a predetermined program stored in the auxiliary storage device 13, thereby achieving the functional configuration processing of the random number seed generation unit 111, the aggregation unit 112, and the model management unit 113. The random number seed generation unit 111 sends a pseudo-random number seed to each client 2. The random number seed is a predetermined number of digits that form the basis of the pseudo-random number. The random number seeds to be sent may be determined randomly, or they may be sent in order from smallest to largest value.

[0033] The aggregation unit 112 receives evaluation values ​​for random number seeds from each client 2 and performs aggregation. The evaluation value for random number seeds can also be described as an evaluation value that represents the performance of a machine learning model when the parameters obtained based on the random number seed are applied to the machine learning model. The method of aggregating the evaluation values ​​may be any method, such as calculating the sum, mean, median, and standard deviation of the evaluation values ​​from each client 2. The method of aggregating the evaluation values ​​from each client 2 is not limited to the predetermined method. The aggregation unit 112 transmits the aggregated evaluation values ​​to each client 2.

[0034] The model management unit 113 updates the machine learning model 121 based, for example, on the aggregated evaluation values ​​obtained by the aggregation unit 112. For example, if the aggregated evaluation values ​​obtained by the aggregation unit 112 are better than the aggregated evaluation values ​​for the random number seed corresponding to the parameter 122 currently applied to the machine learning model 121, the model management unit 113 updates the machine learning model 121 by applying parameters using pseudo-random numbers obtained from the random number seed corresponding to the aggregated evaluation values ​​obtained by the aggregation unit 112.

[0035] Memory 12 holds the machine learning model 121, parameters 122, and evaluation value data 123. The machine learning model 121 is, for example, a neural network, deep learning... The machine learning model 121 may be any of the following: a learning model, a convolutional neural network, a recurrent neural network, etc. The machine learning model 121 is not limited to a model that follows a specific algorithm. The parameters 122 are the values ​​of the parameters currently applied to the machine learning model 121. In the first embodiment, the evaluation value data 123 is, for example, the evaluation value and aggregate value for each client 2 for each random seed.

[0036] Next, client 2 is, for example, a PC, smartphone, tablet terminal, and in-vehicle device. In-vehicle device is, for example, a data communication device, a car navigation system, a drive recorder, and an ECU dedicated to the federated learning system 100. As a hardware configuration, client 2 includes, for example, a processor 21, memory 22, auxiliary storage device 23, and communication unit 24.

[0037] The processor 21, memory 22, auxiliary storage device 23, and communication unit 24 are the same as those of the processor 11, memory 12, auxiliary storage device 13, and communication unit 14, respectively. However, the communication unit 24 may be a circuit that can connect to both wired and wireless networks. Furthermore, if the communication unit 24 is a circuit that can connect to a wireless network, the communication unit 24 may be a wireless communication circuit that supports mobile communication systems such as 5G, 4G, 3G, or 6G, wireless communication systems such as Wi-Fi, or DSRC (Dedicated Short Range Communications), etc. Note that in Figure 2, only the elements related to the processing of the federated learning system 100 are extracted and shown for the hardware configuration of client 2.

[0038] The processor 21 executes a predetermined program stored in the auxiliary storage device 23, thereby achieving the functional configuration processing of the random number generation unit 211, the evaluation unit 212, and the model management unit 213. The random number generation unit 211 receives a random number seed from the server 1 and generates pseudorandom numbers from the random number seed. The generated pseudorandom numbers have the same dimensions as the machine learning model 221 described later. It is also possible to generate and use a number of pseudorandom numbers equal to the dimensions (number) of the input values ​​(or explanatory variables) of the machine learning model 221. The pseudorandom number generation algorithm is not limited to a specific one. Furthermore, in the first embodiment, it is assumed that the same pseudorandom number generation algorithm is used by the server 1 and the multiple clients 2.

[0039] The evaluation unit 212 calculates an evaluation value for the random number seed received from server 1 and sends the evaluation value to server 1. When the index of client 2 is represented as i=1 to M, the evaluation data held by the i-th client 2 is denoted as Di. When all clients 2 use the same random number seed, the random number created from the random number seed received on the tth time (t is a positive integer) is denoted as εt. εt is a value that does not depend on index i. The objective function for evaluating the machine learning model 221 when parameter x is applied to the evaluation data Di is denoted as f(x,Di).

[0040] Let xt represent the parameter currently applied to the machine learning model 221. Let xt+εt represent the parameter obtained by the evaluation unit 212 based on the random number εt. xt+εt means that the parameter xt currently applied to the machine learning model 221 is shifted by the random number εt to obtain the parameter xt+εt. However, the method of obtaining the parameter based on the random number εt is not limited to a specific method. However, in the first embodiment, it is assumed that the method of obtaining the parameter based on the random number εt is common to both the server 1 and each client 2.

[0041] The evaluation unit 212 obtains y(t,i)=f(xt+εt,Di) as an evaluation value for the parameter xt+εt obtained from the random number seed received from server 1.

[0042] For example, if machine learning model 221 is a supervised or unsupervised learning model, The objective function f(x,Di) is a function that calculates the similarity or accuracy between the output values ​​of the machine learning model 221 and the correct values ​​for the evaluation data Di, given the input values. For example, if the machine learning model 221 is a reinforcement learning model, the objective function f(x,Di) is a function that calculates a value indicating how well the output values ​​of the machine learning model 221 meet the criteria for the input evaluation data Di.

[0043] If multiple evaluation data sets exist, the evaluation unit 212 calculates an evaluation value for each evaluation data set and sends the aggregated value of these multiple evaluation values ​​to the server 1. The method for aggregating the evaluation values ​​of multiple evaluation data sets may be any method, such as calculating the mean, median, or standard deviation, and is not limited to a specific method.

[0044] The model management unit 213 manages the machine learning model 221. When the model management unit 213 receives aggregated evaluation values ​​for the random number seed from server 1, it updates the machine learning model 221 based on these aggregated evaluation values. The method for updating the machine learning model 221 by the model management unit 213 is the same as that of the model management unit 113.

[0045] Memory 22 holds the machine learning model 221 and parameters 222. In the first embodiment, machine learning model 221 is the same model as machine learning model 121. In this specification, the same machine learning model means that the algorithms used, dimensions, and parameter values ​​are the same. Parameters 222 are the parameter values ​​currently applied to machine learning model 221.

[0046] The auxiliary storage device 23 holds evaluation data 231. Evaluation data 231 is data acquired by client 2 using predetermined sensors. Evaluation data 231 includes, for example, image data, operation data, temperature data, and speed data. The evaluation data corresponds to the training data used in training a typical machine learning model.

[0047] Note that the hardware and functional configurations of Server 1 and Client 2 shown in Figure 2 are examples only, and Server 1 and Client 2 are not limited to the configurations shown in Figure 2. For example, if Client 2 is a smartphone, it may include a touch panel display, microphone, speaker, GPS (Global Positioning System) receiver, camera, and It is also equipped with various other sensors.

[0048] <Processing flow> Figure 3 shows an example of a processing sequence in the federative learning system 100. For simplicity of explanation, only server 1 and one client 2 are shown in Figure 3. However, other clients 2 perform the same operations.

[0049] In S1, the random number seed generation unit 111 of server 1 obtains a random number seed. In S2, the random number seed generation unit 111 of server 1 sends the random number seed to each client 2. In S3, the random number generation unit 211 of client 2 generates pseudorandom numbers from the random number seed received from server 1. In S4, the evaluation unit 212 of client 2 obtains parameter values ​​using the pseudorandom numbers generated from the random number seed received from server 1 and calculates an evaluation value for the machine learning model 221 when those parameter values ​​are applied. In S5, the evaluation unit 212 of client 2 sends the evaluation value for the random number seed to the processor 11. As a result, server 1 receives the evaluation value for the random number seed sent in S2 from each client 2.

[0050] In S6, the aggregation unit 112 of server 1 aggregates the evaluation values ​​of each client 2 for the random number seed sent in S2. In S7, the aggregation unit 112 of server 1 aggregates the evaluation values ​​of each client 2 Next, send the aggregated evaluation values ​​for each client 2 regarding the random number seed sent via S2.

[0051] In S8, the model management unit 113 of server 1 and the model management unit 213 of each client 2 perform model update processing. In the first embodiment, the model update processing is common to both the model management unit 113 of server 1 and the model management unit 213 of each client 2, so the processing of the model management unit 213 of client 2 will be explained as an example. In OP81, the model management unit 213 determines whether the aggregated evaluation value received from server 1 has improved compared to the aggregated evaluation value of the random number seed corresponding to the parameter values ​​currently applied to the machine learning model 221. The evaluation value of the random number seed corresponding to the parameter values ​​currently applied to the machine learning model 221 may be held by client 2, or it may be sent to each client 2 by server 1 in S7. If a larger evaluation value for a random number seed indicates a higher evaluation of the machine learning model when using that random number seed, then an increase in the aggregated evaluation value received from server 1 indicates improvement. Furthermore, if a smaller evaluation value for a random number seed indicates a higher evaluation of the machine learning model when using that random number seed, then a smaller aggregate evaluation value received from server 1 indicates improvement.

[0052] If the aggregated evaluation values ​​received from server 1 have improved (OP81:YES), the process proceeds to OP82. In OP82, the model management unit 213 of client 2 updates the machine learning model 221 by applying parameters obtained based on the random number seed received from server 1 in S2. The model management unit 213 also overwrites and saves parameter 222 with the parameters obtained based on the random number seed received from server 1 in S2. If the aggregated evaluation values ​​received from server 1 have not improved (OP81:NO), the model management unit 213 does not update the machine learning model 221. Note that the model update process in S8 shown in Figure 3 is just an example, and the model update process is not limited to the process shown in Figure 3.

[0053] In the federated learning system 100, one training session of the machine learning model is terminated, for example, by performing the process shown in Figure 3 for a predetermined number of random number seeds, or when the aggregated evaluation value for the random number seeds exceeds or falls below a predetermined threshold. Furthermore, in the next training session, each x2 will use different evaluation data. Also, in the next training session, a different set of random number seeds may be used.

[0054] <Effects of the First Embodiment> In the first embodiment, when training a machine learning model in a distributed manner across multiple devices, the data flowing between, for example, Server 1 and each Client 2 consists of a random number seed, an evaluation value for the random number seed, and an aggregate value of the evaluation value. All of this data is only a few bytes in size. In conventional federated learning, the machine learning model itself or the parameters of the machine learning model flow between the server and the clients. These are in the megabyte or gigabyte range. Therefore, according to the federated learning system 100 of the first embodiment, when training a machine learning model in a distributed manner across multiple devices, the amount of communication data flowing between Server 1 and each Client 2, and the communication load on Server 1 and each Client 2, can be reduced. This allows for the reduction of the time required to train the machine learning model in the federated learning system 100, even if, for example, some Client 2 are located in areas where low communication speeds are unavailable, resulting in a gap in communication speeds among Client 2. This enables the machine learning model to be trained using data from areas with communication difficulties, such as disaster areas or the digital divide. Furthermore, compression of the transmitted data (random number seed, evaluation value for the random number seed, and aggregate value of the evaluation value) can be applied. Since no processing is performed, the processing load caused by data transmission can be suppressed on server 1 and each client 2.

[0055] In the first embodiment, it is difficult to reconstruct the evaluation data held by each client 2 from the random number seed, the evaluation value for the random number seed, and the aggregated evaluation value that flow between the server 1 and each client 2. Therefore, the federated learning system 100 according to the first embodiment can further enhance the protection of privacy.

[0056] <Second Embodiment> The federated learning system according to the second embodiment is a system consisting only of clients 2, without a server. Figure 4 is a diagram showing an example of the system configuration of the federated learning system 100B according to the second embodiment. The federated learning system 100B includes multiple clients 2. In the federated learning system 100B, one of the multiple clients 2 performs processing such as generating random number seeds and aggregating evaluation values, similar to the server 1 according to the first embodiment. Note that the client 2 that takes on the role of server 1 may change each time learning is performed. The method for selecting the client 2 that takes on the role of server 1 is not limited to a specific method. For example, information such as communication throughput and processor performance may be exchanged among the clients 2, and the client 2 with the best performance may be selected as the client 2 that takes on the role of server 1. In the second embodiment, the client 2 that takes on the role of server 1 is an example of the "first information processing device". That is the case.

[0057] <Other variations> The embodiments described above are merely examples, and this disclosure may be modified as appropriate without departing from its essence.

[0058] In the first embodiment, it was assumed that one random number seed is sent from server 1 to each client 2. However, the invention is not limited to this, and multiple random number seeds may be sent from server 1 to each client 2. In this case, each client 2 obtains an evaluation value for each of the multiple random number seeds received from server 1 and sends it to server 1. Server 1 obtains an aggregate value of the evaluation values ​​from each client 2 for each of the multiple random number seeds. Server 1 sends to each client 2 the random number seed with the best aggregate value of evaluation values ​​among the multiple random number seeds, along with the aggregate value of the evaluation values ​​of that random number seed. For example, server 1 and each client 2 may use the aggregate value of the best evaluation value among the multiple random number seeds to perform model update processing (see S8 in Figure 3), similar to the first embodiment.

[0059] In the first embodiment, it was assumed that a random number seed is sent from server 1 to each client 2. However, the invention is not limited to this, and supplementary information may also be sent from server 1 to each client 2 in addition to the random number seed. Supplementary information specifies the pseudorandom numbers to be used for obtaining parameters. For example, the supplementary information specifies that the pseudorandom numbers to be generated should be obtained after repeating the process 100 times, starting from the random number seed that is sent together with the server. Even if the parameter generation information is information other than the random number seed, both the parameter generation information and the supplementary information may be sent from server 1 to each client 2. Note that the supplementary information is also text data and is only a few bytes in size.

[0060] In the first embodiment, it was described on the premise that Server 1 and the multiple clients 2 have the same machine learning model and the same parameters. However, it is not limited to this, and Server 1 and the multiple clients 2 may have different machine learning models. In this case, even if Server 1 and the multiple clients 2 use the same random number seed, they will generate pseudorandom numbers of different dimensions that match the machine learning models they each have. The Zoom may be different for Server 1 and for each of the multiple Client 2.

[0061] In the first embodiment, the server 1 was described on the premise that it sends the same random number seed to multiple clients 2. However, it is not limited to this, and the server 1 may send different random number seeds to each client 2. For example, when the process shown in Figure 3 is executed once, the server 1 may aggregate the evaluation values ​​it has received from each of the multiple random number seeds sent to each client 2, along with the evaluation values ​​it has received from each client 2 up to that point, and send the random number seed with the best aggregate value at that time and its aggregate value to each client 2. The server 1 and each client 2 update their machine learning model based on the random number seed with the best aggregate value at that time and the aggregate value.

[0062] In addition, the server 1 and multiple clients 2 may each have different machine learning models, and the server 1 may send a different random number seed to each client 2.

[0063] The processes and methods described in this disclosure can be freely combined and implemented, provided that no technical inconsistencies arise.

[0064] Furthermore, a process described as being performed by a single device may be divided and executed by multiple devices. Conversely, a process described as being performed by different devices may be executed by a single device. In a computer system, the hardware configuration (server configuration) by which each function is implemented can be flexibly changed.

[0065] The present disclosure can also be realized by supplying a computer program implementing the functions described in the embodiments above to a computer, and having one or more processors in the computer read and execute the program. Such a computer program may be provided to the computer by a non-temporary computer-readable storage medium that can be connected to the computer's system bus, or it may be provided to the computer via a network. Non-temporary computer-readable storage mediums include, for example, any type of disk such as magnetic disks (floppy disks, hard disk drives (HDDs), etc.), optical disks (CD-ROMs, DVDs, Blu-ray discs, etc.), read-only memory (ROM), random access memory (RAM), EPROM, EEPROM, magnetic cards, flash memory, optical cards, and any type of medium suitable for storing electronic instructions. [Explanation of symbols]

[0066] 1. Server 2. Client 11. Processor 12...memory 13...Auxiliary storage device 14. Communications Department 21 Processor 22...memory 23...Auxiliary storage device 24. Communications Department 100 Associative Learning Systems 111. Random number seed generation unit 112 ··Tallying Department 113 ··Model Management Department 121. Machine Learning Models 122. Parameters 123. Evaluation Value Data 211 ··Random number generation unit 212. Evaluation Department 213 ··Model Management Department 221. Machine Learning Models 222...Parameters 231. Evaluation data

Claims

1. The first information processing device, The first information, used to obtain parameter values ​​for a machine learning model, is transmitted to multiple second information processing devices. The plurality of second information processing devices, The evaluation value of the machine learning model is obtained when the parameter values ​​obtained based on the first information are applied to the machine learning model. The evaluation value is transmitted to the first information processing device. The first information processing device, The multiple evaluation values ​​received from the multiple second information processing devices are aggregated, The aggregated results of the evaluation values ​​are transmitted to the plurality of second information processing devices. The first information processing device and the plurality of second information processing devices Based on the aggregated results of the aforementioned evaluation values, the machine learning model is updated. The first information processing device and the plurality of second information processing devices The machine learning model is updated by setting parameters based on the first piece of information that has the best aggregated evaluation value among multiple pieces of first information. method.

2. The first piece of information is a predetermined number of strings. The method according to claim 1.

3. Transmitting first information used to obtain parameter values ​​for a machine learning model to multiple information processing devices, The plurality of information processing devices receive an evaluation value of the machine learning model when the parameter values ​​obtained based on the first information are applied to the machine learning model, The process involves aggregating the multiple evaluation values ​​received from the multiple information processing devices, The aggregated results of the evaluation values ​​are transmitted to the multiple information processing devices, The machine learning model is updated based on the aggregated results of the aforementioned evaluation values. A control unit that executes Equipped with, The control unit, The machine learning model is updated by setting parameters based on the first piece of information that has the best aggregated evaluation value among multiple pieces of first information. Information processing device.

4. The first information processing device receives first information used to obtain parameter values ​​for a machine learning model, Obtaining an evaluation value of the machine learning model when the parameter values ​​obtained based on the first information are applied to the machine learning model, The evaluation value is transmitted to the first information processing device, The first information processing device receives the aggregated results of multiple evaluation values ​​from multiple information processing devices that have received the first information, The machine learning model is updated based on the aggregated results of the aforementioned evaluation values. A control unit that executes Equipped with, The control unit, The machine learning model is updated by setting parameters based on the first piece of information that has the best aggregated evaluation value among multiple pieces of first information. Information processing device.

Images