Integration device, learning device, and integration method

Abstract

To achieve efficiency of federated learning.SOLUTION: An integration device performs a reception process of receiving a knowledge coefficient relating to first learning data in a first prediction model of a first learning device from the first learning device, a transmission process of transmitting the first prediction model and data relating to the knowledge coefficients of the first learning data received in the reception process respectively to a plurality of second learning devices, and an integration process of generating an integrated prediction model by integrating a model parameter in a second prediction model generated by causing each of the plurality of second learning devices to learn second learning data and the data relating to the knowledge coefficients by the first prediction model, as a result of transmission in the transmission process.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to an integration device, a learning device, and an integration method. [Background technology]

[0002] Machine learning is one of the technologies that realizes AI (Artificial Intelligence). Machine learning technology consists of a learning process and a prediction process. First, the learning process calculates learning parameters so that the error between the predicted value obtained from the input feature vector and the actual value (true value) is minimized. Next, the prediction process calculates a new predicted value from data that was not used in learning (hereafter referred to as test data).

[0003] So far, methods of calculating and computing learning parameters that maximize the prediction accuracy of the predicted value have been devised. For example, a method called perceptron outputs a predicted value based on the input feature vector and the calculation result of a linear combination of a weight vector. Neural networks are also known as multi-perceptrons, and have the ability to solve linearly inseparable problems by stacking multiple perceptrons in multiple layers. Deep learning is a method that introduces new technologies such as dropout into neural networks, and has attracted attention as a method that can achieve high prediction accuracy. Thus, machine learning technologies have been developed with the aim of improving prediction accuracy, and their prediction accuracy is showing capabilities that exceed those of humans.

[0004] When implementing machine learning technology in society, there are issues other than prediction accuracy, such as security, how to update the model after delivery, and limitations on the use of finite resources such as memory.

[0005] One of the security issues is the confidentiality of data. When creating a predictive model using data that includes personal information, such as in the medical or financial fields, it can be difficult to move the data outside the location where it is stored due to its high confidentiality. In general, machine learning can achieve high prediction accuracy by using a large amount of data for learning.

[0006] When learning is performed using only data acquired at one location, the model may be usable only in a very localized area due to the small number of data samples, regional characteristics, etc. In other words, machine learning technology is needed that makes it possible to create a predictive model that achieves high predictions for all of the rich variety of data from each location, without the data being taken out of the location.

[0007] In Non-Patent Document 1, the problem of data confidentiality is overcome by using a federated learning technique. A prediction model is generated by learning each data of each site using one common model as an initial value. The model parameters of the generated prediction model are transmitted to a server, and a process of generating a global prediction model from the model parameters of the prediction model using a coefficient according to the amount of data learned by the server is repeated. Finally, a global prediction model that achieves high prediction accuracy for data of all sites is generated. In addition, Non-Patent Document 2 discloses continuous learning. [Prior art documents] [Patent documents]

[0008] [Patent Document 1] JP 2020-149656 A [Non-patent literature]

[0009] [Non-Patent Document 1] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson and Blaise Aguera y Arcas, “Communication-efficient learning of deep networks from decentralized data”, In Artificial Intelligence and Statistics, pp. 1273-1282, 2017. [Non-Patent Document 2] De Lange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G. and Tuytelaars, T., “Continual learning: A comparative study on how to defy forgetting in classification tasks”, arXiv preprint arXiv:1909.08383 2019. Summary of the Invention [Problem to be solved by the invention]

[0010] In federated learning techniques such as that described in Non-Patent Document 1, the more times the generation of a predictive model at each site and the generation of a global predictive model at the server are repeated, the more time it takes to determine the global predictive model and the amount of communication between the sites and the server increases.

[0011] In addition, when new data is added to the bases or when new bases appear, it is necessary to regenerate the integrated predictive model from the beginning, including the bases that have data that has already been learned. This is because, in machine learning, catastrophic forgetting generally occurs, in which knowledge of previously learned data is lost when new data is learned. In such cases, it is necessary to have a high degree of redundancy in order to relearn data that has already been learned, and to continue storing the data.

[0012] In other words, since data is accumulated on a daily basis, there is a high demand for services using machine learning to update predictive models at any time through continuous learning, as in Non-Patent Document 2, to create predictive models that can respond to new knowledge as well as past knowledge.

[0013] An object of the present invention is to improve the efficiency of associative learning. [Means for solving the problem]

[0014] An integration device according to one aspect of the invention disclosed in the present application is an integration device having a processor that executes a program and a storage device that stores the program, wherein the processor executes a receiving process for receiving from a first learning device a knowledge coefficient related to first learning data in a first prediction model of the first learning device, a transmitting process for transmitting the first prediction model and data related to the knowledge coefficient of the first learning data received by the receiving process to each of a plurality of second learning devices, and an integration process for generating an integrated prediction model by integrating model parameters in a second prediction model generated by each of the plurality of second learning devices having the first prediction model learn the second learning data and the data related to the knowledge coefficient as a result of the transmission process.

[0015] A learning device according to one aspect of the invention disclosed in the present application is a learning device having a processor that executes a program and a storage device that stores the program, wherein the processor executes a learning process of generating a first prediction model by training a learning target model with first learning data; a first transmission process of transmitting model parameters in the first prediction model generated by the learning process to a computer; a reception process of receiving from the computer an integrated prediction model generated by the computer by integrating the model parameters and other model parameters in another first prediction model of another learning device as the learning target model; a knowledge coefficient calculation process of calculating a knowledge coefficient of the first learning data in the first prediction model when the integrated prediction model is received by the reception process; and a second transmission process of transmitting the knowledge coefficient calculated by the knowledge coefficient calculation process to the computer.

[0016] A learning device according to another aspect of the invention disclosed in the present application is a learning device having a processor for executing a program and a storage device for storing the program, wherein the processor executes a first receiving process for receiving from a computer a first integrated prediction model that integrates a plurality of first prediction models and data related to a knowledge coefficient for each of the first learning data used in learning each of the first prediction models; a learning process for training the first integrated prediction model received by the first receiving process using second learning data and the data related to the knowledge coefficients received by the first receiving process as a learning target model to generate a second prediction model; and a transmission process for transmitting model parameters in the second prediction model generated by the learning process to the computer. Effect of the Invention

[0017] According to the representative embodiment of the present invention, it is possible to improve the efficiency of associative learning. Problems, configurations and effects other than those described above will become apparent from the following description of the embodiment. [Brief description of the drawings]

[0018] [Figure 1] FIG. 1 is an explanatory diagram showing an example of associative learning. [Diagram 2] FIG. 2 is a diagram illustrating an example of associative learning that suppresses catastrophic forgetting according to the first embodiment. [Diagram 3] FIG. 3 is a block diagram showing an example of the hardware configuration of a computer. [Figure 4] FIG. 4 is a block diagram of an example of a functional configuration of the computer according to the first embodiment. [Diagram 5] FIG. 5 is a block diagram showing an example of the functional configuration of the learning unit 412. As shown in FIG. [Figure 6] FIG. 6 is a flowchart of an example of an integration process procedure performed by the server according to the first embodiment. [Figure 7] FIG. 7 is a flowchart of an example of a learning process procedure performed by a site according to the first embodiment. [Figure 8] FIG. 8 is a flowchart illustrating a detailed example of a procedure of the first integration process (step S601) by the server illustrated in FIG. [Figure 9] FIG. 9 is a flowchart illustrating a detailed example of a procedure of the second integration process (step S602) by the server illustrated in FIG. [Figure 10] FIG. 10 is a flowchart illustrating a detailed example of the first learning process (step S701) by the base illustrated in FIG. [Figure 11] FIG. 11 is a flowchart illustrating a detailed example of the second learning process (step S702) by the base illustrated in FIG. [Figure 12] FIG. 12 is an explanatory diagram showing a display example 1 of the display screen. [Figure 13] FIG. 13 is an explanatory diagram showing a second display example of the display screen. [Figure 14] FIG. 14 is an explanatory diagram showing a display example 3 of the display screen. [Figure 15] FIG. 15 is an explanatory diagram showing a display example 4 of the display screen. [Figure 16]FIG. 16 is a block diagram of an example of a functional configuration of the server according to the second embodiment. [Figure 17] FIG. 17 is a block diagram of an example of a functional configuration of a base according to the second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0019] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described with reference to the accompanying drawings. In the following, in all the drawings for explaining the embodiments of the present invention, the same reference numerals are used to designate parts having the same functions, and the repeated description thereof will be omitted. EXAMPLES

[0020] <catastrophic oblivion> In general, catastrophic forgetting occurs in machine learning, where knowledge of previously learned training data is lost when current training data is learned. For example, suppose that in phase 1, image data of apples and oranges is trained, and in phase 2, image data of grapes and peaches is trained on a predictive model that can distinguish between images of apples and oranges. Then, the predictive model can distinguish between images of grapes and peaches, but will no longer be able to distinguish between images of apples and oranges.

[0021] As a solution, in Phase 2, a predictive model capable of recognizing images of apples and oranges is trained with image data of apples, oranges, grapes, and peaches, generating a predictive model capable of recognizing images of all four types of images. However, this method requires that the image data of apples and oranges trained in Phase 1 be saved in Phase 2. Also, compared to training using only image data of grapes and peaches in Phase 2, training using image data from both Phase 1 and Phase 2 requires a longer learning time because the amount of data to be trained is greater.

[0022] Catastrophic forgetting that is expected when machine learning technology is implemented in society can be seen in the medical and financial fields. In the field of cancer treatment, treatment methods are evolving rapidly, such as the development of new therapeutic drugs and improvements in proton beam irradiation technology. In order to predict the effectiveness of treatments that keep up with the latest medical technology, it is necessary to update the prediction model to match the evolution of treatments. In the investment field, in order to predict profits and losses that reflect the rapidly changing social situation, it is necessary to update the prediction model by taking into account not only the learning data of the most recent transactions, but also the learning data from many years of the past that have been affected by important factors such as employment statistics and business indicators, as well as natural disasters.

[0023] In particular, in the fields of medicine and finance, when creating a predictive model using training data that includes personal information, it can be difficult to move the training data outside the location where it is stored due to the high level of confidentiality of the training data. One possible solution is to use federated learning.

[0024] Federated learning is a learning method in which a single common prediction model is used as the initial value, and learning is performed using each learning data of each site, generating a prediction model for each site. Federated learning makes it possible to predict both new learning data that occurs over time and learning data that has been learned in the past. Each of the model parameters of the prediction model generated for each site is sent to a server. The server integrates the model parameters of each site to generate an integrated prediction model. By repeating this process, the integrated prediction model achieves the desired prediction accuracy.

[0025] <Associative learning> Fig. 1 is an explanatory diagram showing an example of federated learning. Each of a plurality of locations (four locations 101-104 in Fig. 1 as an example) that serve as learning devices in Fig. 1 holds learning data T1-T4 (when there is no need to distinguish between them, they are simply referred to as learning data T), and is prohibited from sending the learning data T1-T4 outside the locations 101-104.

[0026] The server 100 is an integration device that integrates the prediction models M1 to M4 generated at the bases 101 to 104. The server 100 has a base prediction model (hereinafter, base prediction model) M0. The base prediction model M0 may be an untrained neural network, or may be a trained neural network in which model parameters such as weights and biases are set.

[0027] The bases 101 to 104 are computers that have learning data T1 to T4 and generate prediction models M1 to M4 using the learning data T1 to T4. Each of the learning data T1 to T4 is a combination of learning data and correct answer data as input.

[0028] In phase 1, learning data T1 from site 101 and learning data T2 from site 102 are used, and in phase 2, in addition to the learning data T1 from site 101 and the learning data T2 from site 102 used in phase 1, learning data T3 from site 103 and learning data T4 from site 104 are used.

[0029] [Phase 1] In phase 1, the server 100 transmits the base prediction model M0 to the site 101 and the site 102. The site 101 and the site 102 perform learning using the base prediction model M0 and respective learning data T1 and T2 to generate prediction models M1 and M2, respectively.

[0030] The base 101 and the base 102 each transmit model parameters θ1, θ2 such as weights and biases of the prediction models M1, M2 to the server 100. The server 100 executes an integration process of the received model parameters θ1, θ2 to generate an integrated prediction model M10. The server 100 repeats the update process of the integrated prediction model M10 until the generated integrated prediction model M10 achieves a desired prediction accuracy. Note that the base 101 and the base 102 may transmit the gradients of the model parameters θ1, θ2 of the prediction models M1, M2 to the server 100.

[0031] The integration process is a process of calculating the average value of the model parameters θ1 and θ2. If the number of samples of the learning data T1 and T2 is different, a weighted average may be calculated based on the number of samples of the learning data T1 and T2. In addition, the integration process may be a process of calculating the average value of the gradients of the model parameters θ1 and θ2 transmitted from each of the bases 101 and 102 instead of the model parameters θ1 and θ2.

[0032] The update process of the integrated prediction model M10 is a process in which the server 100 transmits the integrated prediction model M10 to the bases 101 and 102, the bases 101 and 102 input learning data T1 and T2 to the integrated prediction model M10, respectively, perform learning, transmit the model parameters θ1 and θ2 of the regenerated prediction models M1 and M2 to the server 100, and the server 100 regenerates the integrated prediction model M10. When the generated integrated prediction model M10 achieves the desired prediction accuracy, phase 1 ends.

[0033] [Phase 2] In phase 2, the server 100 transmits the integrated prediction model M10 generated in phase 1 to the locations 101 to 104. The locations 101 to 104 respectively input the learning data T1 to T4 to the integrated prediction model M10 to perform learning and generate prediction models M1 to M4. Then, the locations 101 to 104 respectively transmit the model parameters θ1 to θ4 of the generated prediction models M1 to M4 to the server 100. Note that the locations 101 to 104 may transmit the gradients of the model parameters θ1 to θ4 of the prediction models M1 to M4 to the server 100.

[0034] The server 100 executes an integration process of the received model parameters θ1 to θ4 to generate an integrated prediction model M20. The server 100 repeats the update process of the integrated prediction model M20 until the generated integrated prediction model M20 achieves a desired prediction accuracy.

[0035] In the integration process in phase 2, the average value of the model parameters θ1 to θ4 is calculated. If the number of pieces of learning data T1 to T4 is different, a weighted average may be calculated based on the number of pieces of learning data T1 to T4. In addition, the integration process may be a process of calculating the average value of the gradients of the model parameters θ1 to θ4 transmitted from each of the locations 101 to 104 instead of the model parameters θ1 to θ4.

[0036] In the update process of the integrated prediction model M20 in phase 2, the server 100 transmits the integrated prediction model M20 to the locations 101 to 104, the locations 101 to 104 respectively input learning data T1 to T4 to the integrated prediction model M20 to perform learning, and transmit the model parameters θ1 to θ4 of the regenerated prediction models M1 to M4 to the server 100, which then regenerates the integrated prediction model M20. When the generated integrated prediction model M20 achieves the desired prediction accuracy, phase 2 ends.

[0037] If we ignore the repetition of the update process, the transmission and reception between server 100 and points 101 to 104 is 4 times in phase 1 and 8 times (number of arrows) in phase 2, for a total of 12 times. If we take into account the repetition of the update process, an additional 4 times the number of repetitions in phase 1 and 8 times the number of repetitions in phase 2 are required.

[0038] Note that, regarding the prediction accuracy in phase 1 and phase 2, each site calculates it by applying test data other than the learning data T1 to T4 to the integrated prediction models M10, M20. Specifically, for example, if the integrated prediction models M10, M20 are regression models, the prediction accuracy is calculated as the mean square error, the root mean square error, or the coefficient of determination, and if they are classification models, the prediction accuracy is calculated as the accuracy rate, the precision rate, the recall rate, or the F-measure. Alternatively, data for calculating the accuracy of the integrated prediction model stored in the server 100 may be used.

[0039] <Associative learning suppresses catastrophic forgetting> FIG. 2 is an explanatory diagram showing an example of the associative learning that suppresses catastrophic forgetting according to the first embodiment. In FIG. 2, differences from FIG. 1 will be mainly described. Phase 1 is almost the same as the associative learning shown in FIG. 1. The difference from the associative learning shown in FIG. 1 is that when the generated integrated prediction model M10 achieves a desired prediction accuracy, the bases 101 and 102 calculate a knowledge coefficient I1 of the learning data T1 for the prediction model M1 and a knowledge coefficient I2 of the learning data T2 for the prediction model M2, and transmit them to the server 100. The knowledge coefficients I1 and I2 are coefficients of regularization terms that constitute a loss function that accumulates knowledge of the learning data T1 and T2.

[0040] Furthermore, the integrated prediction model M10 may be used to calculate each knowledge coefficient, the prediction model M1 and the integrated prediction model M10 may be used to calculate the knowledge coefficient I1, and the prediction model M2 and the integrated prediction model M10 may be used to calculate the knowledge coefficient I2.

[0041] In phase 2, the server 100 transmits the integrated prediction model M10 and the knowledge coefficients I1 and I2 generated in phase 1 to each of the sites 103 and 104. The sites 103 and 104 input the learning data T3 and T4 to the integrated prediction model M10, respectively, to perform learning, and generate prediction models M3I and M4I that take into account the knowledge coefficients I1 and I2. Then, the sites 103 and 104 transmit the model parameters θ3I and θ4I of the generated prediction models M3I and M4I, respectively, to the server 100. Note that the sites 103 and 104 may transmit the gradients of the model parameters θ3I and θ4I of the prediction models M3I and M4I to the server 100.

[0042] The server 100 executes an integration process of the received model parameters θ3I and θ4I to generate an integrated prediction model M20I. The server 100 repeats the update process of the integrated prediction model M20I until the generated integrated prediction model M20I achieves a desired prediction accuracy.

[0043] In the integration process in phase 2, the average value of the model parameters θ3I and θ4I is calculated. If the number of data in the learning data T3 and T4 is different, a weighted average may be calculated based on the number of data in the learning data T3 and T4. In addition, the integration process may be a process of calculating the average value of each gradient of the model parameters θ3I and θ4I transmitted from each site instead of the model parameters θ3I and θ4I.

[0044] In the update process of the integrated prediction model M20I in phase 2, the server 100 transmits the integrated prediction model M20I to the sites 103 and 104, the sites 103 and 104 respectively input learning data T3 and T4 to the integrated prediction model M20I to perform learning, and transmit the model parameters θ3I and θ4I of the prediction models M3I and M4I regenerated by taking into account the knowledge coefficients I1 and I2 to the server 100, which then regenerates the integrated prediction model M20I. When the generated integrated prediction model M20I achieves the desired prediction accuracy, phase 2 ends.

[0045] Each of the site 103 and the site 104 uses the knowledge coefficient I1 of the learning data T1 of the site 101 and the knowledge coefficient I2 of the learning data T2 of the site 102 during learning. This allows the server 100 to generate an integrated prediction model M20I capable of predicting the learning data T1 of the site 101, the learning data T2 of the site 102, the learning data T3 of the site 103, and the learning data T4 of the site 104, without each of the site 103 and the site 104 using the learning data T1 of the site 101 and the learning data T2 of the site 102 again.

[0046] Ignoring the repetition of the update process, the number of transmissions and receptions between server 100 and locations 101 to 104 is four in phase 1 and four in phase 2 (number of arrows), totaling eight, which is two-thirds of the number in Figure 1.

[0047] Furthermore, when the repetition of the update process is taken into account, four times the number of repetitions in phase 1 and four times the number of repetitions in phase 2 are required. Here too, the total number of transmissions and receptions can be reduced by halving the number of repetitions in phase 2. Furthermore, since the learning data T1 at site 101 and the learning data T2 at site 102 are not used for learning in phase 2, there is no need to store them. Therefore, the capacity of the storage device of server 100 can be used for other processes or data storage, thereby achieving operational efficiency.

[0048] In phase 1, there are bases 101 and 102, but there may be only base 101. In this case, the server 100 does not need to generate the integrated prediction model M10, and it is sufficient to transmit the prediction model M1 and the knowledge coefficient I1, which are the calculation sources of the knowledge coefficient I1, to the bases 103 and 104. Hereinafter, the associative learning that suppresses catastrophic forgetting shown in FIG. 2 will be specifically described.

[0049] <Example of Hardware Configuration of Computers (Server 100, Sites 101 to 104)> FIG. 3 is a block diagram showing an example of a hardware configuration of a computer. The computer 300 has a processor 301, a storage device 302, an input device 303, an output device 304, and a communication interface (communication IF) 305. The processor 301, the storage device 302, the input device 303, the output device 304, and the communication IF 305 are connected by a bus 306. The processor 301 controls the computer 300. The storage device 302 serves as a working area for the processor 301. The storage device 302 is a non-transient or temporary recording medium that stores various programs and data. Examples of the storage device 302 include a ROM (Read Only Memory), a RAM (Random Access Memory), a HDD (Hard Disk Drive), and a flash memory. The input device 303 inputs data. Examples of the input device 303 include a keyboard, a mouse, a touch panel, a numeric keypad, and a scanner. The output device 304 outputs data. The output device 304 may be, for example, a display or a printer. The communication IF 305 connects to a network and transmits and receives data.

[0050] <Example of functional configuration of computer 300> 4 is a block diagram illustrating an example of a functional configuration of a computer 300 according to Example 1. The computer 300 includes a calculation unit 410 including a prediction model integration unit 411 and a learning unit 412, a communication IF 305 including a transmission unit 421 and a reception unit 422, a storage device 302, and an output unit 431.

[0051] 5 is a block diagram showing an example of a functional configuration of the learning unit 412. The learning unit 412 has a knowledge coefficient generating unit 501, a learning unit 502, and a knowledge coefficient synthesizing unit 503. Specifically, the calculation unit 410 and the output unit 431 are realized by, for example, making the processor 301 execute a program stored in the storage device 302 shown in FIG.

[0052] The prediction model integration unit 411 executes integration processing to generate one integrated prediction model M10, M20I based on the model parameters (θ1, θ2), (θ3, θ4) of each prediction model (M1, M2), (M3, M4) transmitted from the multiple locations 101 to 104. For example, a prediction model that has learned a feature vector x in the learning data T is expressed using the model output y, model parameter θ, and function h as shown in the following formula (1).

[0053]

number

[0054] In phase 2, the server 100 receives model parameters θ t For the integrated prediction model M10 composed of the above, the model parameters θ of K prediction models (prediction models M3I and M4I in Fig. 2) generated by learning with K different learning data (T3 and T4 in Fig. 2) at K (K=2 in Fig. 2) bases (bases 103 and 104 in Fig. 2) are k Using the average sum of the gradients gk with respect to t+1 In the following formula (2), η is the learning rate, N is the total number of samples of all the learning data used for learning at K locations (T3, T4 in Figure 2), and Nk is the number of samples of the data used for learning at location k.

[0055]

number

[0056] Here, in the above formula (2), the model parameters θ of the prediction models (in FIG. 2, prediction models M3I and M4I) generated by learning k different learning data Tk at k locations are kThe gradient gk with respect to the model parameters θ3I and θ4I in Fig. 2 is used, but this is a security-conscious method to prevent the training data (T3 and T4 in Fig. 2) from being analyzed. k In addition, prediction models M3I and M4I may be integrated in a manner different from that of formula (2) above, depending on the structure of the prediction model (prediction models M3I and M4I in FIG. 2) such as a fully connected layer or a convolutional layer, the design of the loss function, and the like.

[0057] The learning unit 412 starts from a prediction model composed of model parameters determined by random initial values ​​or a base prediction model M0, and generates a prediction model by learning using the learning data T, and synthesizes knowledge coefficients in the knowledge coefficient synthesis unit 503. The learning unit 412 also generates a prediction model by learning using the synthesized knowledge coefficients synthesized in the knowledge coefficient synthesis unit 503 and the learning data T.

[0058] Specifically, for example, when the computer 300 is at the locations 101 and 102, the learning unit 412 acquires the base prediction model M0 from the server 100, learns using the learning data T1, generates the prediction model M1, and generates the knowledge coefficient I1 in the knowledge coefficient generation unit 501. Similarly, for the location 102, a prediction model M2 is generated using the learning data T2, and the knowledge coefficient generation unit 501 generates the knowledge coefficient I2.

[0059] Furthermore, when the computer 300 is the location 103, the learning unit 412 acquires the knowledge coefficients I1 and I2 of the locations 101 and 102 from the server 100 and then synthesizes them in the knowledge coefficient synthesis unit 503. Similarly, for the location 104, when the learning unit 412 acquires the knowledge coefficients I1 and I2 of the locations 101 and 102 from the server 100, then synthesizes them in the knowledge coefficient synthesis unit 503. Furthermore, at the locations 103 and 104, the knowledge coefficient generation unit 501 may generate knowledge coefficients I3 and I4 in preparation for future increases in the number of locations.

[0060] Also, in the locations 103 and 104, the learning unit 412 may generate a prediction model M3I using the combined knowledge coefficient generated by the knowledge coefficient combination unit 503 of the server 100 and the learning data T3 of the location 103. Similarly, for the location 104, the learning unit 412 generates a prediction model M4I using the combined knowledge coefficient combined by the knowledge coefficient combination unit 503 of the server 100 and the learning data T4 of the location 104.

[0061] The learning unit 502 uses the above formula (1) to obtain the input learning data T m Feature vector x m The predicted value y obtained from m and the correct label t, which is the actual value or class number. m The model parameters θ m Calculate the loss function L(θ m ) is set. m is a number that identifies the learning data T.

[0062] Specifically, for example, the learning unit 502 selects the past learning data T to be considered among the knowledge coefficients for the past learning data T generated by the knowledge coefficient generating unit 501. m Regarding the past knowledge term R(θ m ) to set the

[0063] Loss function L(θ m ) is expressed by the error function E(θ m ) and past knowledge term R(θ m ) and

[0064]

number

[0065] Past knowledge term R(θ m ) is, for example, as shown in the following formula (4), a coefficient λ of the regularization term and a combined knowledge coefficient Ω generated by the knowledge coefficient combining unit 503. ij and the model parameters θ obtained by learning mand the model parameters θ of the base prediction model M0 B Here, i and j indicate the j-th unit in the i-th layer in the prediction model M.

[0066]

number

[0067] The knowledge coefficient generation unit 501 extracts knowledge of the learning data T by calculating a knowledge coefficient I using the learning data T and a prediction model M that is learned and generated using the learning data T. Specifically, for example, there is a method of extracting knowledge by using the knowledge coefficient I in a regularization term.

[0068] As shown in the following formula (5), the knowledge coefficient I ij (x m ;θ m ) is the training data T m The model parameters θ m The model parameters θ of the output of the prediction model M ij The training data T m Knowledge Coefficient I ij (x m ;θ m ) is the training data T m and the training data T m Since the prediction model M is generated using only the learning data T and the prediction model M (for example, the learning data T1 and T2 and the prediction models M1 and M2 in FIG. 2), it is not necessary to store the past learning data T and the prediction model M. In addition, the past learning data T and the prediction model M can be used to store the learning data T in the future. m Knowledge Coefficient I ij (x m ;θ m ) and the training data T m The learning data T in the future from the time of learning m+1 The model parameters θ m+1 The knowledge coefficient I is generated using ij (x m ;θ m+1) and there is no need to store them in order to generate them.

[0069]

number

[0070] The knowledge coefficient synthesis unit 503 synthesizes a plurality of knowledge coefficients generated using the learning data T to be introduced from among the knowledge coefficient group generated by the knowledge coefficient generation unit 501, and generates a synthetic knowledge coefficient. Specifically, for example, the knowledge coefficient synthesis unit 503 of the server 100 or the bases 103 and 104 synthesizes a plurality of knowledge coefficients I1 and I2 generated using the learning data T1 and T2, and generates a synthetic knowledge coefficient Ω(I1, I2).

[0071] Further, the knowledge coefficient synthesis unit 503 synthesizes the learning data T based on U, which stores the identification number of the knowledge coefficient I to be introduced, as shown in the following formula (6). m Feature vector x m In the sample p direction in , the sum of the knowledge coefficients I to be introduced is calculated, and normalization is performed by the total number of samples. In this example, a method is used to introduce and store knowledge of specific data using an L2 norm type regularization term, but it is also possible to use an L1 norm type or an elastic net, or to use knowledge that has been stored by converting data such as a replay-based method or a parameter isolation-based method, or to add knowledge to the base prediction model M0 using the learning data T m or a network path may be used.

[0072]

number

[0073] The transmission unit 421 transmits various data. Specifically, for example, if the computer 300 is the server 100, the transmission unit 421 transmits the base prediction model M0 and the first integrated prediction model M10 to the locations 101 and 102 during learning at each location (phase 1). In addition, the transmission unit 421 transmits the integrated prediction models M10 and M20I generated by the prediction model integration unit and the knowledge coefficients I1 and I2 (or the combined knowledge coefficient Ω(I1, I2)) to the locations 103 and 104 during learning at each location (phase 2). In addition, the transmission unit 421 transmits to each location whether to continue or end the repetition of the federated learning based on the accuracy verification result executed at each location.

[0074] In addition, if the computer 300 is a location 101, 102, the transmission unit 421 transmits to the server 100 the learned model parameters θ1, θ2 during learning (phase 1) at each location 101, 102, all knowledge coefficients I1, I2 so far or knowledge coefficients I1, I2 input by the operator to be used for learning at each location 101, 102, and the accuracy verification results of the prediction models M1, M2.

[0075] Also, if the computer 300 is at the location 103, 104, the transmission unit 421 transmits to the server 100 the learned model parameters θ3I, θ4I and the accuracy verification results of the prediction models M3I, M4I during learning at each of the locations 103, 104 (phase 2).

[0076] The receiving unit 422 receives various data. Specifically, for example, if the computer 300 is the server 100, during the prediction model integration (phase 1), the receiving unit 422 receives the model parameters θ1, θ2, the knowledge coefficients I1, I2, and the prediction accuracy verification results of the prediction models M1, M2 from the locations 101, 102. Furthermore, during the prediction model integration (phase 2), the receiving unit 422 receives the model parameters θ3I, θ4I, and the accuracy verification results of the prediction models M3I, M4I from the locations 103, 104.

[0077] Moreover, if the computer 300 is at the locations 101 and 102, the receiving unit 422 receives the base prediction model M0 and the first integrated prediction model M10 during learning (phase 1) at each of the locations 101 and 102. Moreover, if the computer 300 is at the locations 103 and 104, the receiving unit 422 receives the integrated prediction models M10 and M20I and the knowledge coefficients I1 and I2 (or the composite knowledge coefficient Ω) during learning (phase 2) at each of the locations 103 and 104.

[0078] In addition, the data transmitted and received is converted by encryption or the like from the viewpoint of security. This makes it difficult to analyze the data used for learning from the prediction model M.

[0079] <Example of integration process> 6 is a flowchart showing an example of an integration process procedure by the server 100 according to the first embodiment. The server 100 judges whether or not to send the knowledge coefficient I to the base (step S600). If the knowledge coefficient I is not sent to the base (step S600: No), this means the start of phase 1. Therefore, the server 100 executes a first integration process for integrating a plurality of prediction models M1 and M2 (step S601).

[0080] On the other hand, when the knowledge coefficient I is sent to the base (step S600: Yes), phase 1 is completed. Therefore, the server 100 executes a second integration process to integrate a plurality of prediction models M3 and M4 (step S602). Details of the first integration process (step S601) will be described later in FIG. 8, and details of the second integration process (step S602) will be described later in FIG. 9. Even if the knowledge coefficient I is not sent, an identification code indicating phase 1 or phase 2 may be sent together with the base prediction model M0 or the integrated prediction model used as the base prediction model M0, and it may be determined whether to execute step S601 or step S602 based on the identification code.

[0081] <Example of learning process procedure> 7 is a flowchart showing an example of a learning process procedure by a base according to the first embodiment. The base judges whether or not the base has received the knowledge coefficient I from the server 100 (step S700). If the base has not received the knowledge coefficient I (step S700: No), the base is a base (e.g., bases 101 and 102) that performs learning without using the knowledge coefficient I. Therefore, the bases 101 and 102 execute a first learning process (step S701).

[0082] On the other hand, if the knowledge coefficient I is received (step S700: Yes), the location is a location (for example, the locations 103 and 104) that performs federated learning using the knowledge coefficient I. The locations 103 and 104 execute a second learning process (step S702). Details of the first learning process (step S701) will be described later in FIG. 10, and details of the second learning process (step S702) will be described later in FIG. 11. Even if the knowledge coefficient is not received, an identification code indicating whether it is phase 1 or phase 2 may be received together with the base prediction model M0 or the integrated prediction model M used as the base prediction model M0, and it may be determined whether to execute step S701 or step S702 based on the received code.

[0083] <First integration process (step S601)> FIG. 8 is a flowchart showing a detailed processing procedure example of the first integration process (step S601) by the server 100 shown in FIG. 6. The server 100 sets a transmission target model to the bases 101 and 102 determined as the transmission destination in step S600: No (step S801). Specifically, for example, if the base prediction model M0 has not yet been transmitted, the server 100 sets the base prediction model M0 as the transmission target, and if an instruction is given at the time of transmission target model setting in step S801 to set the integrated prediction model M10 that has been transmitted in the past and was generated at that time as the base prediction model, the integrated prediction model M10 is set as the transmission target. In the latter case, since the knowledge coefficient I related to the past knowledge of the data learned in the past is not transmitted together, the knowledge of the data learned in the past is forgotten in the newly generated prediction model M. Then, the server 100 transmits the transmission target model to each of the bases 101 and 102 (step S802).

[0084] Next, the server 100 receives the model parameters θ1 and θ2 of the prediction models M1 and M2 from each of the bases 101 and 102 (step S803). Then, the server 100 generates an integrated prediction model M10 using the received model parameters θ1 and θ2 (step S804). Then, the server 100 transmits the generated integrated prediction model M10 to each of the bases 101 and 102 (step S805).

[0085] Next, the server 100 receives the prediction accuracy by the integrated prediction model M10 from each of the bases 101, 102 (step S806). Then, the server 100 verifies each prediction accuracy (step S808). Specifically, for example, the server 100 determines whether each prediction accuracy is equal to or greater than a threshold. Note that the prediction accuracy by the integrated prediction model M10 for the data of each of the bases 101, 102 is calculated at each base, but if the server 100 has evaluation data, the prediction accuracy by the integrated prediction model M10 for the evaluation data may be used. After this, the server 100 transmits the verification result to each of the bases 101, 102 (step S808).

[0086] The server 100 judges whether or not the overall prediction accuracy is equal to or greater than the threshold in the verification result (step S809). If the overall prediction accuracy is not equal to or greater than the threshold (step S809: No), that is, if there is even one prediction accuracy less than the threshold, the process returns to step S803, and the server 100 waits for the model parameters θ1 and θ2 of the prediction models M1 and M2 updated again from the bases 101 and 102.

[0087] On the other hand, if the overall prediction accuracy is equal to or greater than the threshold (step S809: Yes), each of the bases 101, 102 calculates and transmits knowledge coefficients I1, I2 for the integrated prediction model M10, and the server 100 receives the knowledge coefficients I1, I2 for the integrated prediction model M10 from each of the bases 101, 102 (step S810). Then, the server 100 stores the integrated prediction model M10 and the knowledge coefficients I1, I2 in the storage device 302 (step S811). This ends the first integration process (step S601).

[0088] <Second integration process (step S602)> Fig. 9 is a flowchart showing a detailed example of the processing procedure of the second integration process (step S602) by the server 100 shown in Fig. 6. In step S600: Yes, the server 100 sets a transmission target model and knowledge coefficients to the locations 103 and 104 determined as the transmission destination (step S901). The server 100 transmits the integrated prediction model M10 and knowledge coefficients I1 and I2 to the locations 103 and 104 determined as the transmission destination (step S902). Note that the knowledge coefficient may be a composite knowledge coefficient Ω generated in advance by the server 100.

[0089] Next, the server 100 receives the model parameters θ3I and θ4I of the integrated prediction model M10 (or the integrated prediction model M20I if the integrated prediction model M20I has already been received) from each of the bases 103 and 104 (step S903). Then, the server 100 generates the integrated prediction model M20I using the received model parameters θ3I and θ4I (step S904). Then, the server 100 transmits the generated integrated prediction model M20I to each of the bases 103 and 104 (step S905).

[0090] Next, the server 100 receives the prediction accuracy by the integrated prediction model M201 from each of the bases 103 and 104 (step S906). Then, the server 100 verifies each prediction accuracy (step S907). Specifically, for example, the server 100 judges whether each prediction accuracy is equal to or higher than a threshold. Note that the prediction accuracy by the integrated prediction model M201 for the data of each of the bases 103 and 104 is calculated at each base, but if the server has evaluation data, the prediction accuracy by the integrated prediction model M201 for the evaluation data may be used. After this, the server 100 transmits the verification result to each of the bases 103 and 104 (step S908).

[0091] The server 100 judges whether or not the overall prediction accuracy is equal to or greater than the threshold in the verification result (step S909). If the overall prediction accuracy is not equal to or greater than the threshold (step S909: No), that is, if there is even one prediction accuracy less than the threshold, the process returns to step S903, and the server 100 waits for the model parameters θ3I and θ4I of the prediction model M20I updated again from each of the bases 103 and 104.

[0092] On the other hand, if the overall prediction accuracy is equal to or greater than the threshold (step S909: Yes), each of the bases 103, 104 calculates and transmits knowledge coefficients I3, I4 for the integrated prediction model M201, and the server 100 receives the knowledge coefficients I3, I4 for the integrated prediction model M201 from each of the bases 103, 104 (step S910). Then, the server 100 stores the integrated prediction model M20I and the knowledge coefficients I3, I4 in the storage device 302 (step S911). This ends the second integration process (step S602).

[0093] <First Learning Process (Step S701)> Fig. 10 is a flowchart showing a detailed example of the processing procedure of the first learning process (step S701) by the bases 101 and 102 shown in Fig. 7. In step S700: No, each of the bases 101 and 102 stores the base prediction model M0 from the server 100 in the storage device 302 (step S1001). Note that when the base prediction model M0 is the integrated prediction model M10, the knowledge coefficient I related to the past knowledge of the data learned in the past is not transmitted together, and therefore the knowledge of the data learned in the past is forgotten in the newly generated prediction model M.

[0094] Next, each of the bases 101 and 102 learns the base prediction model M0 using the learning data T1 and T2 to generate prediction models M1 and M2 (step S1002). Then, each of the bases 101 and 102 transmits the model parameters θ1 and θ2 of the prediction models M1 and M2 to the server 100 (step S1003). As a result, the server 100 generates an integrated prediction model M10 (step S804).

[0095] Thereafter, each of the bases 101 and 102 receives the integrated prediction model M10 from the server 100 (step S1004). Then, each of the bases 101 and 102 calculates the prediction accuracy of the integrated prediction model M10 (step S1005) and transmits it to the server 100 (step S1006). As a result, the server 100 executes verification of each prediction accuracy (step S807).

[0096] After that, each of the bases 101 and 102 receives the verification result from the server 100 (step S1007). Then, each of the bases 101 and 102 judges whether the total prediction accuracy is equal to or greater than the threshold in the verification result (step S1008). If the total prediction accuracy is not equal to or greater than the threshold (step S1008: No), that is, if there is even one prediction accuracy less than the threshold, each of the bases 101 and 102 re-learns the integrated prediction model M10 as a base prediction model using the learning data T1 and T2 (step S1009), and transmits the model parameters θ1 and θ2 of the prediction models M1 and M2 generated by re-learning to the server 100 (step S1010). Then, returning to step S1004, each of the bases 101 and 102 waits for the integrated prediction model M10 from the server 100.

[0097] On the other hand, if the overall prediction accuracy is equal to or greater than the threshold (step S1008: Yes), the bases 101 and 102 calculate the knowledge coefficients I1 and I2 for the prediction models M1 and M2 (step S1011) and transmit them to the server 100 (step S1012). This ends the first learning process (step S701).

[0098] <Second Learning Process (Step S702)> Fig. 11 is a flowchart showing a detailed example of the processing procedure of the second learning process (step S702) by the locations 101 and 102 shown in Fig. 7. Each of the locations 103 and 104 that has transitioned to step S700: Yes stores the integrated prediction model M10 and the knowledge coefficients I1 and I2 from the server 100 in the storage device 302 (step S1101).

[0099] Next, each of the bases 103 and 104 synthesizes the learning data T3 and T4 and the knowledge coefficients I1 and I2 to generate a synthetic knowledge coefficient Ω (step S1102), learns the integrated prediction model M10 using the synthetic knowledge coefficient Ω, and generates prediction models M3I and M4I (step S1103). Note that when the knowledge coefficient is received from the server 100 as a synthetic knowledge coefficient Ω that has been generated in advance, it is not necessary to perform step S1102 of generating a synthetic knowledge coefficient from the knowledge coefficient I at the base.

[0100] Then, each of the bases 103 and 104 transmits the model parameters θ3I and θ4I of the prediction models M3I and M4I to the server 100 (step S1104). As a result, the server 100 generates an integrated prediction model M20I (step S904).

[0101] Next, each of the bases 103 and 104 receives the integrated prediction model M20I from the server 100 (step S1105). Then, each of the bases 103 and 104 calculates the prediction accuracy of the integrated prediction model M20I (step S1106) and transmits it to the server 100 (step S1107). As a result, the server 100 executes verification of each prediction accuracy (step S907).

[0102] After that, each of the bases 103 and 104 receives the verification result from the server 100 (step S1108). Then, each of the bases 103 and 104 judges whether the total prediction accuracy is equal to or greater than the threshold in the verification result (step S1109). If the total prediction accuracy is not equal to or greater than the threshold (step S1109: No), that is, if there is even one prediction accuracy less than the threshold, each of the bases 103 and 104 combines the knowledge coefficients I1 and I2 to generate a combined knowledge coefficient Ω (step S1110). The combined knowledge coefficient Ω generated in step S1102 may be temporarily stored in a memory and used.

[0103] Then, each of the bases 103 and 104 re-learns the prediction model M20I as a base prediction model using the learning data T3 and T4 and the synthetic knowledge coefficient Ω (step S1110), and transmits the model parameters θ3I and θ4I of the prediction models M3I and M4I generated by re-learning to the server 100 (step S1111). Then, returning to step S1105, each of the bases 103 and 104 waits for the updated integrated prediction model M20I again from the server 100.

[0104] On the other hand, if the overall prediction accuracy is equal to or greater than the threshold (step S1109: Yes), the bases 103 and 104 calculate the knowledge coefficients I3 and I4 for the prediction models M3 and M4 (step S1112) and transmit them to the server 100 (step S1113). This ends the second learning process (step S702).

[0105] In this way, according to the above-described learning system, it is possible to generate a prediction model M20 capable of predicting the learning data T1-T4 at the multiple locations 101-104 without moving the learning data T1-T4 at the multiple locations 101-104 outside the locations, and by using the knowledge coefficients I1, I2 of the multiple learning data T1, T2 learned in the past, without using the learning data T1, T2 learned in the past again for learning. It is possible to generate an integrated prediction model M20I capable of predicting the learning data T1-T4 at the multiple locations 101-104, which is generated by repeating learning at each of the locations 103, 104 and model integration at the server 100.

[0106] By applying continuous learning technology to the integrated prediction model M20I at the locations 103 and 104, it is possible to generate a prediction model capable of predicting the learning data T1 to T4 at the locations 101 to 104 by using the learning data T3 and T4 and the knowledge coefficients I1 and I2 of the multiple learning data T1 and T2 learned in the past, without using the learning data T1 and T2 learned in the past for re-learning. This makes it possible to generate a prediction model M20 capable of predicting the learning data T1 to T4 at the locations 101 to 104.

[0107] <Display screen example> Next, an example of a display screen displayed on a display, which is an example of the output device 304 of the computer 300, or on a display of the computer 300 that is the output destination from the output unit 431, will be described.

[0108] 12 is an explanatory diagram showing a display example 1 of a display screen. A display screen 1200 is displayed on the displays of the bases 103 and 104, for example.

[0109] The display screen 1200 includes a Select train data button 1201 , a Select knowledge button 1202 , a Train button 1203 , a mode name field 1204 , a data name field 1205 , the selection screen 810 , and check boxes 811 .

[0110] When a user at each of the bases 103 and 104 wishes to conduct training, the user selects “Train” in the mode name column 1204. Next, the user at each of the bases 103 and 104 presses the Select train data button 1201 to select training data T3 and T4. The selected training data T3 and T4 are displayed in the data name column 1205.

[0111] Furthermore, the user at each of the bases 103, 104 selects a knowledge coefficient indicating past knowledge to be incorporated into the prediction model, for example, by checking a check box 1211. The knowledge coefficient synthesis unit 503 at each of the bases 103, 104 synthesizes the checked knowledge coefficients I1, I2. The synthesized knowledge coefficient Ω generated by synthesis is used in learning when the user at each of the bases 103, 104 presses the Train button 1203 (step S1103). Note that the knowledge coefficient to be selected may be presented or determined in advance at the request of the server 100.

[0112] 13 is an explanatory diagram showing a display example 2 of the display screen. The display screen 1300 is a screen that is displayed when the server 100 generates an integrated prediction model. The display screen 1300 includes a Select client button 1301, a Start button 1302, a mode name field 1204, a data name field 1205, a selection screen 1310, and a check box 1311.

[0113] When the user of the server 100 wishes to generate a predictive model that integrates predictive models, the user selects “Federation” in the mode name field 1204. Next, the user of the server 100 presses a Select client button 1301 and selects the location for generating the integrated predictive model, for example by checking a check box 1311.

[0114] The prediction model integration unit 411 of the server 100 integrates prediction models from bases having the checked client names using formula (2) (steps S804, S904). Note that in the selection screen 1310, for example, a "1" may be displayed in the Train query field for a base that has alerted the server 100 that new learning data to be learned has been accumulated, or for a base that has transmitted the latest base prediction model M0. After that, by pressing the Start button 1302, prediction models are generated and integrated, and an integrated prediction model is generated (steps S804, S904).

[0115] FIG. 14 is an explanatory diagram showing a display example 3 of the display screen. The display screen 1400 is a screen for confirming the prediction accuracy in the server 100. Specifically, for example, the server 100 first learns with one learning data T1. After that, the site 101 learns with the learning data T2 using the knowledge coefficient I1 when learning with the learning data T1, and the site 102 learns with the learning data T3 using the knowledge coefficient I1 when learning with the learning data T1. The server 100 integrates the prediction model learned by the site 101 with the learning data T2 and the prediction model learned by the site 102 with the learning data T3. The display screen 1400 is a result display example when the number of repetitions when the integration process is performed is "1". Specifically, the display screen 1400 is displayed when the server 100 performs prediction accuracy verification (step S907) to determine whether the prediction accuracy at the sites 101 and 102 is equal to or higher than a threshold (step S909).

[0116] The display screen 1400 includes a View results button 1401 , a View status button 1402 , a mode name column 1204 , a data name column 1205 , an associative learning result display screen 1411 , and a data status screen 1412 .

[0117] When the user of the server 100 wishes to check the prediction accuracy of the integrated prediction model, he / she selects "Federation" in the mode name column 1204. When the federated learning process instructed in Fig. 13 has ended or the prediction accuracy is being verified (step S807, step S907), a View results button 1401 and a View status button 1402 are displayed. When the View results button 1401 is pressed, the prediction accuracy of the integrated prediction model based on each of the learning data T1 to T3 is displayed as shown in a federated learning result display screen 1411.

[0118] When the View status button 1402 is pressed, a list is displayed on a data status screen 1412 showing at which locations the learning data T1 to T3 were obtained and learned.

[0119] As shown on the associative learning result display screen 1411, in the integrated prediction model generated by associative learning of a prediction model trained on the learning data T2 of the site 101 and a prediction model trained on the learning data T3 of the site 102 using the knowledge coefficient I1 of the learning data T1 previously trained on the server 100, it can be seen that not only the prediction accuracy based on the learning data T2 of the site 101 (P(T2) = 92.19%) and the prediction accuracy based on the learning data T3 of the site 102 (P(T3) = 94.39%) are high, but also the prediction accuracy based on the learning data T1 previously trained on the server 100 (P(T1) = 98.44%) is high.

[0120] Fig. 15 is an explanatory diagram showing a display example 4 of a display screen. A display screen 1500 is a screen that displays results related to a prediction model on the server 100. Specifically, for example, similar to the case of Fig. 14, the server 100 first learns with one learning data T1. Thereafter, the site 101 learns with learning data T2 using the knowledge coefficient I1 when learning with the learning data T1, and the site 102 learns with learning data T3 using the knowledge coefficient I1 when learning with the learning data T1. The server 100 integrates the prediction model learned by the site 101 with the learning data T2 and the prediction model learned by the site 102 with the learning data T3.

[0121] Furthermore, in FIG. 15, the server 100 displays the results of the integrated prediction model generated by learning new learning data T4 of the server 100 for the integrated prediction model using the knowledge coefficient I1 when learning with the learning data T1, the knowledge coefficient I2 of the learning data T2 for the integrated prediction model, and the knowledge coefficient I3 of the learning data T3.

[0122] The display screen 1500 includes a View results button 1401 , a View status button 1402 , a mode name field 1204 , a data name field 1205 , a learning result screen 1511 , and a data status screen 1412 .

[0123] When a user of the server 100 wants to check the prediction accuracy of the prediction model, he / she selects “Train” in the mode name column 1204. When the learning process instructed in FIG. 12 has ended, a View results button 1401 and a View status button 1402 are displayed.

[0124] Pressing the View results button 1401 displays the prediction accuracy of each piece of training data by the final prediction model, as in a training result screen 1511. Pressing the View status button 1402 displays a list of the location at which each piece of training data was obtained and the training data, as in a data status screen 1412.

[0125] As displayed on the learning result screen 1511, an integrated prediction model generated by joint learning of a prediction model trained on learning data T2 at site 101 and a prediction model trained on learning data T3 at site 102 using a knowledge coefficient I1 of learning data T1 previously trained on the server 100 is set as a base prediction model M0.

[0126] Furthermore, a prediction model M4 is generated by continuous learning using the base prediction model M0, the learning data T4, the knowledge coefficient I1 of the learning data T1, the knowledge coefficient I2 of the learning data T2, and the knowledge coefficient I3 of the learning data T3. In this case, it can be seen that not only the prediction accuracy based on the learning data T2 of the site 101 (P(T2)=91.84%) and the prediction accuracy based on the learning data T3 of the site 102 (P(T3)=92.15%), but also the prediction accuracy based on the learning data T1 previously learned by the server 100 (P(T1)=98.27%) and the prediction accuracy based on the learning data T4 of the server 100 learned this time (P(T4)=96.31%) can be maintained high.

[0127] In the first embodiment, the prediction models M1, M2, M3I, and M4I that are the targets of federated learning are generated only at the locations 101 to 104. However, the prediction models generated at the server 100 may be the targets of federated learning. Also, any of the locations 101 to 104 may play the role of the server 100.

[0128] Furthermore, the bases 101-104 may generate a prediction model without using the knowledge coefficient I of the past learning data T. In this case, the bases 101-104 generate a prediction model by learning using the knowledge coefficient I at the bases that have generated a prediction model that has passed the verification result from the server 100 (prediction accuracy is equal to or greater than a threshold). Then, the server 100 may integrate the prediction models generated at a limited number of the bases 101-104 based on the verification result to generate a final integrated prediction model. Note that the bases may be divided into groups in advance based on the distribution characteristics of the data, rather than the verification result, and an integrated prediction model may be generated for each group.

[0129] 15, a prediction model M20 capable of predicting the learning data T1-T4 at the multiple locations 101-104 can be generated without moving the learning data T1-T4 at the multiple locations 101-104 outside the locations, and without using the learning data T1, T2 learned in the past for re-learning by using the knowledge coefficients I1, I2 of the learning data T1, T2. An integrated prediction model M20I capable of predicting the learning data T1-T3 at the multiple locations 101-103 can be generated by repeating learning at each of the locations 103, 104 and model integration at the server 100.

[0130] When continuous learning technology is applied to the integrated prediction model M20I at the location 104, it is possible to generate a prediction model capable of predicting the learning data T1 to T4 at the multiple locations 101 to 104 by using the learning data T4 and the knowledge coefficients I1, I2, and I3 of the multiple previously learned learning data T1, T2, and T3, without using the previously learned learning data T1, T2, and T3 again for learning. This makes it possible to generate a prediction model M20 capable of predicting the learning data T1 to T4 at the locations 101 to 104.

[0131] Therefore, it is possible to shorten the update time for the prediction model by reducing the amount of learning data, reduce the amount of communication traffic by reducing the number of communicating locations and the number of communications, and reduce the usage of the storage device 302 by eliminating the need to store past data.

[0132] In addition, in the first embodiment, since each computer 300 has the prediction model integration unit 411 and the learning unit 412, each computer 300 can execute as the server 100 and the locations 101 to 104. In addition, in the first embodiment, the number of locations in phase 1 is two, but the number of locations in phase 1 may be three or more. Similarly, the number of locations in phase 2 is two, but the number of locations in phase 2 may be three or more.

[0133] After the locations 101 to 104 transmit the knowledge coefficients I1 to I4 to the server 100, the locations 101 to 104 no longer need the learning data T1 to T4. Therefore, the locations 101 to 104 may delete the learning data T1 to T4. This allows the storage devices 302 of the locations 101 to 104 to consume less memory. EXAMPLES

[0134] A second embodiment will be described. In the second embodiment, the server 100 and the locations 101 to 104 have their respective roles unified, and the device configuration is minimized, compared to the first embodiment. The server 100 does not generate a prediction model using learning data. The locations 101 to 104 do not integrate the prediction models. Note that the same components as those in the first embodiment are denoted by the same reference numerals, and the description thereof will be omitted.

[0135] 16 is a block diagram illustrating an example of a functional configuration of the server 100 according to the embodiment 2. Compared to FIG.

[0136] Fig. 17 is a block diagram illustrating an example of a functional configuration of the bases according to the embodiment 2. Compared to Fig. 4, the bases 101 to 104 do not include the prediction model integrating unit 411.

[0137] In this way, according to the second embodiment, as in the first embodiment, it is possible to achieve a reduction in the update time of the prediction model by reducing the amount of learning data, a reduction in the amount of communication traffic by reducing the number of communicating locations and the number of communications, and a reduction in the usage of the storage device 302 by eliminating the need to store past data.

[0138] The present invention is not limited to the above-described embodiments, and includes various modified examples and equivalent configurations within the spirit of the appended claims. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to those having all of the configurations described. Also, a part of the configuration of one embodiment may be replaced with a configuration of another embodiment. Also, a configuration of another embodiment may be added to a configuration of one embodiment. Also, a part of the configuration of each embodiment may be added, deleted, or replaced with another configuration.

[0139] In addition, each of the above-mentioned configurations, functions, processing units, processing means, etc. may be realized in hardware, for example by designing some or all of them as an integrated circuit, or may be realized in software by a processor interpreting and executing a program that realizes each function.

[0140] Information such as programs, tables, files, etc. that realize each function can be stored in a storage device such as a memory, a hard disk, or an SSD (Solid State Drive), or in a recording medium such as an IC (Integrated Circuit) card, an SD card, or a DVD (Digital Versatile Disc).

[0141] In addition, the control lines and information lines shown are those considered necessary for the explanation, and do not necessarily show all the control lines and information lines necessary for implementation. In reality, it can be considered that almost all components are connected to each other. [Explanation of symbols]

[0142] 100 Server (integrated device) 101~104 Bases (Learning Devices) 300 Computers 301 Processor 302 Storage Devices 410 Arithmetic unit 411 Prediction Model Integration Department 412 Learning Department 421 Transmitter 422 Receiving unit 431 Output section 501 Knowledge Coefficient Generation Unit 502 Learning Department 503 Knowledge Coefficient Synthesis Unit

Claims

1. An integrated device having a processor for executing a program and a storage device for storing the program, The processor, a receiving process for receiving, from a first learning device, a knowledge coefficient related to first learning data in a first prediction model of the first learning device; a transmission process for transmitting the first prediction model and data related to a knowledge coefficient of the first learning data received by the reception process to each of a plurality of second learning devices; an integration process for generating an integrated prediction model by integrating model parameters in a second prediction model generated by each of the second learning devices having the first prediction model learn the second learning data and the data related to the knowledge coefficient as a result of the transmission process; An integrated device comprising:

2. 2. The integrated device of claim 1, A device capable of communicating with a plurality of first learning devices; The processor, executing a pre-integration process for generating a pre-integrated prediction model by integrating model parameters in a first prediction model generated by each of the plurality of first learning devices by training a first learning object model using the first learning data; In the transmission process, the processor receives a knowledge coefficient related to the first learning data from each of the plurality of first learning devices; In the transmission process, the processor transmits, to each of the second learning devices, the advance integration prediction model generated by the advance integration process and data related to a knowledge coefficient for each of the first learning data received by the reception process; In the integration process, the processor integrates model parameters in a second prediction model generated by each of the second learning devices having a second learning target model learn the second learning data and the data related to the knowledge coefficient as a result of transmission in the transmission process, thereby generating the integrated prediction model.

16. An integrated device comprising:

3. 3. The integrated device of claim 2, In the advance integration process, the processor repeats a process of generating the advance integrated prediction model and transmitting the generated model as the first learning target model to each of the first learning devices until the prediction accuracy of each of the first prediction models becomes equal to or greater than a first threshold value.

16. An integrated device comprising:

4. 3. The integrated device of claim 2, In the receiving process, when the prediction accuracy of each of the plurality of first prediction models becomes equal to or greater than a first threshold value, the processor receives a knowledge coefficient related to the first learning data from each of the plurality of first learning devices.

16. An integrated device comprising:

5. 2. The integrated device of claim 1, In the transmission process, the processor transmits the first prediction model and a knowledge coefficient of the first learning data to each of the plurality of second learning devices.

16. An integrated device comprising:

6. 3. The integrated device of claim 2, The processor, executing a synthesis process for synthesizing the knowledge coefficients for each of the first learning data to generate a synthetic knowledge coefficient; In the transmission process, the processor transmits the advance integrated prediction model and the combined knowledge coefficient combined by the combination process to each of the second learning devices.

16. An integrated device comprising:

7. 3. The integrated device of claim 2, In the transmission process, the processor repeats a process of generating the integrated prediction model and transmitting the integrated prediction model as the second learning target model to each of the second learning devices until the prediction accuracy of each of the second prediction models becomes equal to or greater than a second threshold value.

16. An integrated device comprising:

8. A learning device having a processor that executes a program and a storage device that stores the program, The processor, a learning process for generating a first prediction model by training a learning object model using the first training data; a first transmission process of transmitting model parameters in the first prediction model generated by the learning process to a computer; a receiving process of receiving, from the computer, an integrated prediction model generated by the computer by integrating the model parameters and other model parameters in other first prediction models of other learning devices, as the learning target model; a knowledge coefficient calculation process for calculating a knowledge coefficient of the first learning data when the integrated prediction model is received by the receiving process; a second transmission process of transmitting the knowledge coefficient calculated by the knowledge coefficient calculation process to the computer; A learning device that executes the above.

9. The learning device according to claim 8, In the learning process, the processor repeats a process of learning the first learning data into the integrated prediction model to generate the first prediction model until the integrated prediction model is no longer received by the receiving process. A learning device characterized by:

10. The learning device according to claim 8, The processor, executing a prediction accuracy calculation process of calculating a prediction accuracy of a first prediction model generated by training the integrated prediction model with the first training data by the learning process; In the knowledge coefficient calculation process, the processor calculates the knowledge coefficient in the first prediction model when the prediction accuracy calculated by the prediction accuracy calculation process and the prediction accuracy calculated by another learning device become equal to or greater than a first threshold value. A learning device characterized by:

11. A learning device having a processor that executes a program and a storage device that stores the program, The processor, a first receiving process for receiving, from a computer, a first prediction model and data relating to a knowledge coefficient of the first learning data used to learn the first prediction model; a learning process for generating a second prediction model by learning the second learning data and the data related to the knowledge coefficient received by the first receiving process as a learning target model into the first prediction model received by the first receiving process; a transmission process of transmitting model parameters in the second prediction model generated by the learning process to the computer; A learning device that executes the above.

12. The learning device according to claim 11, executes a second receiving process to receive from the computer, as the learning target model, a second integrated prediction model generated by the computer by integrating model parameters in the second prediction model and other model parameters in another second prediction model trained by another learning device; In the learning process, the processor repeats the process of generating the second prediction model until the second integrated prediction model is no longer received from the computer by the second receiving process. A learning device characterized by:

13. The learning device according to claim 11, The processor, a second receiving process for receiving, from the computer, a second integrated prediction model generated by integrating model parameters in the second prediction model and other model parameters in another second prediction model trained by another learning device, as the learning target model; a prediction accuracy calculation process for calculating a prediction accuracy of a second prediction model generated by causing a second integrated prediction model received by the second receiving process to learn the second learning data and the data related to the knowledge coefficient through the learning process; In the learning process, the processor repeats the process of generating the second prediction model until the prediction accuracy calculated in the prediction accuracy calculation process and the prediction accuracy calculated by the other learning device become equal to or greater than a second threshold value. A learning device characterized by:

14. The learning device according to claim 11, In the first reception process, the processor receives from the computer a first integrated prediction model obtained by integrating a plurality of the first prediction models and data on a knowledge coefficient for each of the first learning data used in learning each of the first prediction models; The processor, executing a synthesis process for synthesizing the knowledge coefficients for each of the first learning data to generate a synthetic knowledge coefficient; In the learning process, the processor causes the learning target model to learn the second learning data and the synthesis knowledge coefficient generated by the synthesis process, thereby generating a second prediction model. A learning device characterized by:

15. An integration method using an integration device having a processor that executes a program and a storage device that stores the program, comprising: The processor, a receiving process for receiving, from a first learning device, a knowledge coefficient related to first learning data in a first prediction model of the first learning device; a transmission process for transmitting the first prediction model and data related to a knowledge coefficient of the first learning data received by the reception process to each of a plurality of second learning devices; an integration process for generating an integrated prediction model by integrating model parameters in a second prediction model generated by each of the second learning devices having the first prediction model learn the second learning data and the data related to the knowledge coefficient as a result of the transmission process; 2. An integration method comprising:

Images