Evaluation device, evaluation method, and program
The evaluation device in federated learning uses control coefficients to efficiently assess participant contributions, addressing the computational costs of conventional methods and enhancing the learning process by improving convergence speed and timely reward determination.
Patent Information
- Application Number
- PCT/JP2023/019486
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-25
- Publication Date
- 2024-11-28
AI Technical Summary
Conventional methods for evaluating the contribution of participants in federated learning are computationally costly, making it inefficient to assess and reward their contributions accurately.
An evaluation device that uses control coefficients to assess the degree of contribution by updating aggregated model parameters and local model parameters, allowing for efficient evaluation of participant contributions without incurring large calculation costs.
Enables the evaluation of participant contributions in federated learning efficiently, improving the convergence speed of aggregated model parameters and enabling timely reward determination based on accurate contribution assessment.
Smart Images

Figure JP2023019486_28112024_PF_FP_ABST
Abstract
Description
Evaluation device, evaluation method, and program
[0001] The present invention relates to a technique for evaluating learning participants in a federated learning technique.
[0002] In federated learning, participants (workers) do not share their local learning data, but instead share only the learned model while learning. This allows for the generation of a model that reflects the characteristics of each local learning data while protecting the local learning data.
[0003] Participants in federated learning contribute to accurate learning by providing their local learning data and computational resources. In federated learning, it is important to appropriately reward participants who contribute significantly to learning, and this must be properly controlled.
[0004] Conventionally, the Sharpley Value, learning accuracy, etc. have been calculated to evaluate the degree of contribution to associative learning (see, for example, Non-Patent Document 1). In addition, various methods for evaluating the degree of contribution are known, such as evaluation methods based on game theory and auction methods.
[0005] Rongfei Zeng, Chao Zeng, Xingwei Wang, Bo Li, and Xiaowen Chu, "A Comprehensive Survey of Incentive Mechanism for Federated Learning," CoRR, abs / 2106.15406, 2021, [Retrieved May 12, 2023], Internet: https: / / arxiv.org / abs / 2106.15406
[0006] However, conventional methods have the problem that the computational cost for evaluating the contribution to federated learning is high.
[0007] In view of this, the present invention provides a technique for evaluating the degree of contribution to associative learning without incurring a large computational cost.
[0008] The evaluation device receives a value corresponding to the control coefficient or a value corresponding to a new control coefficient obtained by the associative learning, and obtains an evaluation value corresponding to the model learning device based on the control coefficient or the new control coefficient. Here, the associative learning repeats a learning process in which the aggregate model parameters are updated to local model parameters of each model learning device based on the aggregate model parameters, the local learning data of each of the multiple model learning devices, and the control coefficients of each model learning device, an aggregation process in which the aggregate model parameters are updated to new aggregate model parameters using the local model parameters, and an update process in which the control coefficients are updated to new control coefficients based on the difference between the aggregate model parameters or the new aggregate model parameters and the local model parameters or the new local model parameters.
[0009] As described above, in the present invention, an evaluation value is obtained based on the control coefficient or a new control coefficient, so that the degree of contribution to associative learning can be evaluated without incurring a large computational cost.
[0010] FIG. 1 is a block diagram illustrating the configuration of a federated learning system according to each embodiment. FIG. 2 is a block diagram illustrating the functional configuration of a model learning device according to the first embodiment. FIG. 3 is a block diagram illustrating the functional configuration of an aggregation device according to the first embodiment. FIG. 4 is a flow diagram illustrating the processing of the model learning device according to the first embodiment. FIG. 5 is a flow diagram illustrating the processing of the aggregation device according to the first embodiment. FIG. 6 is a block diagram illustrating the functional configuration of a model learning device according to a second embodiment. FIG. 7 is a block diagram illustrating the functional configuration of the aggregation device according to the second embodiment. FIG. 8 is a flow diagram illustrating the processing of the model learning device according to the second embodiment. FIG. 9 is a flow diagram illustrating the processing of the aggregation device according to the second embodiment. FIG. 10 is a block diagram illustrating the hardware configuration of each device according to the embodiments.
[0011] Hereinafter, embodiments of the present invention will be described with reference to the drawings. [Principle] First, the principle of the embodiment will be described. An evaluation device according to the embodiment receives a value corresponding to a control coefficient or a value corresponding to a new control coefficient obtained through associative learning, and obtains an evaluation value corresponding to a model learning device based on the control coefficient or the new control coefficient. This associative learning repeats the following steps: a learning process for updating the aggregate model parameters to local model parameters of each model learning device based on aggregate model parameters, local learning data of each model learning device, and the control coefficients of each model learning device; an aggregation process for updating the aggregate model parameters to new aggregate model parameters using the local model parameters; and an update process for updating the control coefficients to new control coefficients based on the difference between the aggregate model parameters or the new aggregate model parameters and the local model parameters or the new local model parameters.
[0012] In this associative learning, the control coefficient updated as described above is used in the learning process of each model learning device to correct the difference between the direction of overall learning of the aggregate model parameter and the direction of learning of each model learning device. This reduces the variance of the local model parameters and the aggregate model parameters, thereby improving the convergence speed of the aggregate model parameters. That is, the control coefficient is updated based on the difference between the aggregate model parameter or new aggregate model parameter and the local model parameter or new local model parameter. This corrects the difference between the direction of learning of the aggregate model parameter and the direction of learning of each local model parameter by each model learning device, reduces the variance of the local model parameter, and improves the convergence speed of the aggregate model parameter. Such a control coefficient is based on the change in the difference between the aggregate model parameter or new aggregate model parameter and the local model parameter or new local model parameter. This can also be considered an index representing the contribution of each model learning device to associative learning. For example, in a model learning device with a small control coefficient, the direction of learning of the local model parameter is closer to the direction of learning of the aggregate model parameter, and therefore has a greater impact on the learning of the aggregate model parameter. On the other hand, for example, in a model learning device with a large control coefficient, the direction in which learning of the local model parameters progresses is significantly different from the direction in which learning of the aggregate model parameters progresses. Such a model learning device can be said to have a significant influence on the appropriate updating (correction) of the aggregate model parameters, except in cases where there is malicious intent to disrupt learning. Therefore, the evaluation value obtained based on the control coefficient or a new control coefficient can be said to represent the contribution of each model learning device to the associative learning. Furthermore, the control coefficient is updated for the above-mentioned associative learning, and the evaluation value is obtained by reusing this coefficient. Therefore, in this embodiment, the contribution to the associative learning can be evaluated without incurring significant computational costs.
[0013] For example, the learning process is performed by each model learning device, and the aggregation process is performed by the aggregation device. The aggregation device may be a device other than the model learning device, or one of the model learning devices may also serve as the aggregation device. The update process may be performed by the aggregation device, one of the model learning devices, or a device other than the model learning device or the aggregation device. The device that performs the update process is referred to as the evaluation device. That is, the aggregation device may also serve as the evaluation device, or one of the model learning devices may also serve as the evaluation device, or a dedicated evaluation device other than the aggregation device and the model learning device may be provided. For example, if the aggregation device is a device other than the model learning device, the aggregation device may also serve as the evaluation device, or an evaluation device other than the aggregation device may be provided. For example, if one of the model learning devices also serves as the aggregation device, this model learning device may also serve as the evaluation device, or an evaluation device other than the model learning device may be provided.
[0014] The local model parameters and aggregate model parameters represent model parameters of a specific machine learning model. The model parameters may be columns (e.g., vectors) of coefficients (e.g., real numbers) or scalars (e.g., real numbers), depending on the type of model. Note that there are no limitations on the models used in the present invention. Examples of models include neural networks used in deep learning, hidden Markov models, support vector machines, etc. For example, if the model is a neural network, the local model parameters and aggregate model parameters are columns (vectors) representing weights and biases.
[0015] The evaluation value corresponding to each model learning device is a value based on the above-mentioned control coefficient or new control coefficient. For example, a model learning device with a small control coefficient has a greater influence on learning of aggregate model parameters, so the evaluation value corresponding to such a model learning device may be set to a large value. On the other hand, a model learning device with a large control coefficient has a greater influence on appropriate updating (correction) of aggregate model parameters, so the evaluation value corresponding to such a model learning device may be set to a large value. The evaluation value may also be determined based on both of these criteria, or based on at least one of these criteria and another criterion. That is, for example, when the index based on the control coefficient or the new control coefficient is a first value, the evaluation value is a first evaluation value; when the index is a second value, the evaluation value is a second evaluation value; and when the index is a third value, the evaluation value is a third evaluation value; the first value is smaller than the second value, the second value is smaller than the third value; the first evaluation value represents a greater contribution to associative learning than the second evaluation value, and / or the third evaluation value represents a greater contribution to associative learning than the second evaluation value. A larger evaluation value may indicate a greater contribution to associative learning, or a smaller evaluation value may indicate a greater contribution to associative learning. This index is, for example, a value representing the magnitude of the control coefficient or the new control coefficient. For example, this index is the norm of the control coefficient or the new control coefficient (e.g., L1 norm or L2 norm). Alternatively, this index may be a relative value of the information amount of the control coefficient of each model learning device with respect to a statistical value of the information amount of the control coefficients of multiple model learning devices. Examples of statistical values include the mean, median, maximum value, minimum value, and representative value.
[0016] Preferably, the evaluation device of the embodiment obtains an evaluation value corresponding to the model learning device based on the control coefficients obtained in the learning round process or new control coefficients each time a learning round process including a learning process, an aggregation process, and an update process is performed. As described above, in this embodiment, the contribution to the federated learning can be evaluated without incurring significant computational costs, and therefore the evaluation value can be obtained each time a learning round process is performed. This allows, for example, each model learning device to select local learning data to be used for the next learning round and to select whether or not to participate in the next learning round after each learning round.
[0017] A reward to be given to each model learning device may be determined based on the evaluation value. For example, the evaluation device of the embodiment may (1) select a model learning device to which a reward is to be given from among the multiple model learning devices, or (2) determine a reward value according to the evaluation value, or may perform both of these, based on the evaluation value.
[0018] [First Embodiment] Next, a first embodiment will be described. In this embodiment, an evaluation method for performing federated learning as described in Reference 1 will be exemplified. However, this does not limit the present invention. Reference 1: Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J. Reddi, Sebastian U. Stich, Ananda Theertha Suresh, "SCAFFOLD: Stochastic Controlled Averaging for Federated Learning," Proceedings of the 37th International Conference on Machine Learning, PMLR 119:5132-5143, 2020.
[0019] 1, the federated learning system 1 of this embodiment has model learning devices 11-1, ..., 11-n and an aggregation device 12 (evaluation device), which are configured to be able to communicate over a network. Here, n is an integer greater than or equal to 1, for example, n is an integer greater than or equal to 2. Furthermore, i = 1, ..., n, and each model learning device is represented as model learning device 11-i.
[0020] 2, the model learning device 11-i of this embodiment includes a memory 111-i, a control unit 112-i, a storage unit 113-i, a communication unit 114-i, a learning unit 115-i, and an update unit 116-i. The storage unit 113-i stores local learning data ξ i (r),k where r is an integer between 0 and R-1 (r=0, ..., R-1) that represents the entire learning round. Here, R is an integer between 1 and K-1 that represents the upper limit of the learning round r. k is an integer between 0 and K-1 that represents the local update step executed by the model learning device 11-i for each learning round r. Here, K is an integer between 1 and K-1 that represents the upper limit of the local update step k. Local learning data ξ i (r),k may vary depending on r, may be the same regardless of r, may vary depending on k, or may be the same regardless of k. i (r),k may be determined in advance or may be determined according to input information or processing. Furthermore, model learning device 11-i may participate in all learning rounds or only some of the learning rounds.
[0021] 3, the aggregation device 12 of this embodiment has a memory 121, a control unit 122, a communication unit 124, an aggregation unit 125, an evaluation value calculation unit 126, and a reward determination unit 127. The aggregation device 12 executes each process under the control of the control unit 122. Information input to the aggregation device 12 and information obtained by each unit is stored in the memory 121 one by one, and is read out as needed and used in each process.
[0022] <Processing> Next, an example of the federated learning process of this embodiment will be described. <Processing of Model Learning Device 11-i (FIG. 4)> Step S111-i: As shown in FIG. 4, the communication unit 114-i of the model learning device 11-i (where i∈{1, . . . , n}) (FIG. 2) participating in the federated learning receives the aggregated model parameters transmitted from the aggregating device 12 (FIG. 3). ~ x (r) Initial value of ~ x (0) (r=0) and the control coefficient of the aggregation device 12 ~ c (r) (also called "control value") initial value ~ c (0) (r=0) and the pair ( ~ x (0) , ~ c (0) ) is received. ~ x (r) may be, for example, a sequence of real numbers (vector) or a real number (scalar). ~ c (r) is, for example, a real number (scalar). ~ x (0) , ~ c (0) is predetermined, and any initial value ~ x (0) , ~ c (0) These sets ( ~ x (0) , ~ c (0) ) is sent to the learning unit 115-i. ~ The "~" in "x" should actually be written directly above the "x" as follows: However, due to limitations on writing notation, "~" is sometimes written to the upper left of the "x". ~ The "~" in "c" should be written directly above the "c", but due to limitations on writing notation, the "~" is sometimes written to the upper left of the "c". ~ x (0) , ~ c (0) ) is sent to the learning unit 115-i.
[0023] Step S112-i: The learning unit 115-i calculates the initial value of the aggregate model parameter. ~ x (0) The local model parameters x of the model learning device 11-i i (r) The initial value x i (0) (r=0). That is, the learning unit 115-i sets x i (0) = ~ x (0) Set to.
[0024] Step S113-i: The learning unit 115-i executes the following loop processing of steps S1131-i, S1132-i, S1133-i, S1134-i, S1135-i, and S1136-i for r = 0, ..., R-1. The learning unit 115-i first executes the loop processing for learning round r = 0, and then repeats the process of executing the loop processing for learning round r with r + 1 as the new r each time the loop processing for learning round r is executed, until the loop processing for learning round r = R-1 is completed.
[0025] Step S1131-i: The learning unit 115-i calculates the aggregate model parameters ~ x (r) the local model parameters x of the local update step k of the learning round r i (r),k The initial value x i (r),0 (k=0). That is, the learning unit 115-i sets x i (r),0 = ~ x (r) Set to.
[0026] Step S1132-i: The learning unit 115-i executes the following loop processing of steps S1132a-i and S1132b-i for k = 0, ..., K-1. The learning unit 115-i first executes the loop processing of local update step k = 0, and then repeats the process of executing the loop processing of local update step k with k + 1 as the new k each time the loop processing of local update step k is executed, until the loop processing of local update step k = K-1 is completed.
[0027] Step S1132a-i: The learning unit 115-i calculates the local model parameter x i (r),k , the control coefficient of the aggregation device 12 ~ c (r) , the control coefficient c of the model learning device 11-i i (r) , the local learning data ξ read from the storage unit 113-i i (r),k Using the following calculation, the update amount v i (r),k+1 get. Here, F i represents the loss function in the model learning device 11-i. is the local model parameter x i (r),k The local model is determined by the local training data ξ i (r),k The loss function F when i represents the probability gradient (e.g., unbiased probability gradient) of i (r) The initial value of c i (0) is a vector or real number consisting of real numbers greater than or equal to 0. The initial value c i (0) may be fixed or may be determined depending on the start process or other processes.
[0028] Step S1132b-i: The learning unit 115-i calculates the local model parameter x i (r),k , update amount v i (r),k+1, and a constant η, the following calculation is performed to obtain the local model parameters x i (r),k+1 get. Here, the constant η is a real number greater than 0.
[0029] Step S1133-i: The learning unit 115-i calculates x i (r),K the new local model parameters x i (r+1) That is, the learning unit 115-i sets x i (r+1) = x i (r),K In this way, the learning unit 115-i sets at least the aggregate model parameter ~ x (r) , the local learning data ξ of the model learning device 11-i i (r),k , and the control coefficient c of the model learning device 11-i i (r) Based on the aggregate model parameters ~ x (r) The local model parameters x of the model learning device 11-i i (r+1) The learning unit 115-i executes a learning process to update the new local model parameters x i (r+1) to the update unit 116-i and the communication unit 114-i. Furthermore, the learning unit 115-i sends the aggregate model parameter ~ x (r) and control coefficient ~ c (r) , c i (r) to the update unit 116-i.
[0030] Step S1134-i: The communication unit 114-i receives the new local model parameter x i (r+1) and the new control coefficient c i (r+1) The pair (x i (r+1) , c i (r+1) ) to the aggregation device 12 via the network.
[0031] Step S1135-i The communication unit 114-i receives the new aggregation model parameters transmitted from the aggregation device 12 via the network. ~ x (r+1) and the control coefficient of the aggregation device 12 ~ c (r+1) With the pair ( ~ x (r+1) , ~ c (r+1) ) are received. ~ x (r+1) , ~ c (r+1) ) is sent to the learning unit 115-i. ~ x (r+1) , ~ c (r+1) The generation of the .
[0032] Step S1136-i: The update unit 116-i updates the aggregate model parameters ~ x (r) , new local model parameters x i (r+1) , and control coefficients ~ c (r) , c i (r) The following calculation is performed using the new control coefficient c i (r+1) and output it. Here, K η is a positive real constant. That is, the update unit 116-i updates at least the aggregate model parameters ~ x (r) and the new local model parameters x i (r+1) Based on the difference between i (r) The new control coefficient c i (r+1) The update unit 116-i executes an update process to update the new control coefficient c i (r+1) to the communication unit 114-i and the learning unit 115-i.
[0033] <Processing of Aggregation Device 12 (FIG. 5)> Step S121: As illustrated in FIG. 5, the communication unit 124 of the aggregation device 12 (FIG. 3) receives the above-mentioned aggregation model parameters ~ x (r) Initial value of ~ x (0) (r=0) and the control coefficient of the aggregation device 12 ~ c (r) (also called "control value") initial value ~ c (0) (r=0) and the pair ( ~ x (0) , ~ c (0) ) is transmitted via the network to model learning device 11-j participating in federated learning. Here, j∈N and N ⊆ {1, ..., n}. If all model learning devices 11-1, ..., 11-N participate in federated learning, N = {1, ..., n}; otherwise, N ⊂ {1, ..., n}.
[0034] Step S122: The aggregation device 12 executes the following loop processing of steps S1221, S1222, S1223, S1224, S1225, and S1226 for r = 0, ..., R-1. The aggregation device 12 first executes the loop processing of learning round r = 0, and then repeats the processing of executing the loop processing of learning round r with r + 1 as a new r each time the loop processing of learning round r is executed, until the loop processing of learning round r = R-1 is completed.
[0035] Step S1221: The communication unit 124 of the aggregation device 12 (FIG. 3) receives the local model parameters x transmitted as described above via the network from the model learning device 12-j participating in the federated learning. j (r+1) and control coefficient c j (r+1) The pair (x j (r+1) , c j (r+1) ) are received. j (r+1) , c j (r+1) ) is sent to the aggregation unit 125.
[0036] Step S1222: The aggregating unit 125 aggregates the sent local model parameters x j (r+1) (where j∈N), and calculate the aggregate model parameters in training round r+1 by ~ x (r+1) get. Here, |N| represents the number of elements of N, and represents the number of model learning devices 11-j participating in the federated learning. That is, the aggregator 125 aggregates the local model parameters x j (r+1) Aggregated model parameters using ~ x (r) the new aggregated model parameters ~ x (r+1) Execute aggregation processing to update to.
[0037] Step S1223: The aggregation unit 125 calculates the control coefficient c j (r+1) Using the above, the control coefficient of the aggregation device 12 in the learning round r+1 is calculated as follows: ~ c (r+1) That is, the aggregation unit 125 obtains the control coefficient c j (r+1) The control coefficient of the aggregation device 12 is ~ c (r) is the new control coefficient ~ c (r+1) Execute aggregation processing to update to. The aggregation unit 125 calculates the new control coefficient ~ c (r+1) to the communication unit 124. The aggregation unit 125 also transmits the control coefficient c j (r+1) to the evaluation value calculation unit 126. The aggregation unit 125 further calculates the control coefficient ~ c (r+1) may be sent to the evaluation value calculation unit 126.
[0038] Step S1224: Aggregation model parameters ~ x (r+1) and control coefficient ~ c (r+1) With the pair ( ~ x (r+1) , ~ c (r+1)) is input to the communication unit 125. The communication unit 125 receives the set ( ~ x (r+1) , ~ c (r+1) ) to the model learning device 11-i. For example, the communication unit 125 transmits the set ( ~ x (r+1) , ~ c (r+1) ) may be transmitted only to model learning device 11-j that participated in the federated learning. ~ x (r+1) , ~ c (r+1) ) may be transmitted.
[0039] Step S1225: The evaluation value calculation unit 126 calculates the evaluation value e corresponding to the model learning device 11-i based on the control coefficient. j (r+1) For example, the evaluation value calculation unit 126 calculates and outputs the control coefficient c j (r+1) (new control coefficient), the evaluation value e corresponding to the model learning device 11-i is calculated. j (r+1) Alternatively, the control coefficient c j (r+1) and ~ c (r+1) (new control coefficient) to obtain the evaluation value e j (r+1) For example, the evaluation value e j (r+1) is a scalar. j (r+1) The larger the evaluation value e, the greater the contribution to the federated learning. j (r+1) It may be expressed that the smaller the value, the greater the contribution to associative learning.
[0040] As mentioned above, the control coefficient c j (r+1) The model learning device 11-j with a small aggregate model parameter ~ x (r+1) Since the influence of the model learning device 11-j on the learning is large, the evaluation value e j(r+1) On the other hand, the control coefficient c j (r+1) The model learning device 11-j with a large value of ~ x (r+1) Since the influence on the appropriate update (correction) of the model learning device 11-j is large, the evaluation value e j (r+1) Alternatively, the evaluation value e may be set to a large value based on both of these criteria. j (r+1) may be determined, and the evaluation value e may be determined based on at least one of these criteria and other criteria. j (r+1) may be determined.
[0041] For example, the evaluation value calculation unit 126 calculates the control coefficient c j (r+1) Indicator based on j (r+1) Depending on the j (r+1) may be determined. j (r+1) = e1 for ind j (r+1) = v1 e j (r+1) = e2 for ind j (r+1) = v2 e j (r+1) = e3 for ind j (r+1) = v3 where v1<v2<v3. For example, the evaluation value e j (r+1) = e1 is the evaluation value e j (r+1) = e2. For example, the evaluation value e j (r+1) = e3 is the evaluation value e j (r+1) = e2. For example, the evaluation value e j (r+1) = e1 is the evaluation value e j (r+1) = e2, and the evaluation value ej (r+1) = e3 is the evaluation value e j (r+1) For example, the evaluation value e j (r+1) In the case where the larger the evaluation value e is, the greater the contribution to associative learning, e1 may be greater than e2, e3 may be greater than e2, or e1 may be greater than e2 and e3 may be greater than e2. For example, j (r+1) In the case where the smaller the control coefficient c is, the greater the contribution to associative learning, e1<e2 may be satisfied, or e3<e2 may be satisfied, or e1<e2 and e3<e2 may be satisfied. j (r+1) Indicator based on j (r+1) When v1 (first value) is the evaluation value e j (r+1) is e1 (first evaluation value), and the index ind j (r+1) When v2 (second value) is the evaluation value e j (r+1) is e2 (second evaluation value), and the index ind j (r+1) When v3 (third value) is the evaluation value e j (r+1) is e3 (third evaluation value). v1 (first value) is smaller than v2 (second value), and v2 (second value) is smaller than v3 (third value). e1 (first evaluation value) represents a greater contribution to associative learning than e2 (second evaluation value), and / or e3 (third evaluation value) represents a greater contribution to associative learning than e2 (second evaluation value).
[0042] Control coefficient c j (r+1) Indicator based on j (r+1) For example, j (r+1) The norm of c (for example, L1 norm or L2 norm) may be used. j (r+1) / ~ c (r+1) or the norm of c j (r+1) -~ c (r+1 Alternatively, the control coefficient c j (r+1) Indicator based on j (r+1) is the control coefficient c of the model learning device 11-i for the statistical value of the amount of information of the control coefficients of the plurality of model learning devices. j (r+1) For example, the index ind j (r+1) but, ~ c (r+1) The amount of information I ( ~ c (r+1) ) for c j (r+1) The amount of information I(c j (r+1) ) ratio I(c j (r+1) ) / I( ~ c (r+1) ) or the difference abs(I( ~ c (r+1) )-I(c j (r+1) )) or its function value. Here, abs(α) represents the absolute value of α. Other examples include the index ind j (r+1) is the amount of information I(c 1 (r+1) ), ..., I(c N (r+1) ) is the average, median, maximum, minimum, or representative value of stat, j (r+1) ) ratio I(c j (r+1) ) / stat, or the difference abs(stat-I(c j (r+1) )) or a function value thereof. j (r+1) is sent to the remuneration determination unit 127.
[0043] Step S1226: The reward determination unit 127 calculates the evaluation value e j (r+1) The reward A for the model learning device 11-j isj (r+1) (incentive) is determined and output. j (r+1) The reward may be a monetary value such as electronic money, cryptocurrency, cash, or stocks, or may be a right such as points, usage authority, usage time, usage capacity, or privileges. Examples of methods for determining the reward are given below.
[0044] <Example 1 of Reward Determination Method> The reward determination unit 127 determines the evaluation value e j (r+1) Based on this, the reward A j (r+1) For example, the reward determination unit 127 may select a model learning device 11-i (where i∈{1, ..., N}) that gives the evaluation value e j (r+1) Y model learning devices 11-i are selected in descending order of their contribution to the associative learning based on the j (r+1) and the other model learning devices 11-i′ (where i′∈{1, . . . , N}) are given reward A j (r+1) Here, Y is an integer equal to or greater than 1 and less than N. For example, the evaluation value e j (r+1) indicates a greater contribution to associative learning, the reward determination unit 127 determines the evaluation value e j (r+1) Y model learning devices 11-i are selected in descending order of the evaluation value e j (r+1) In this case, the smaller the evaluation value e, the greater the contribution to the associative learning. j (r+1) Y model learning devices 11-i are selected in ascending order of reward A. j (r+1) may be something whose value varies depending on its size (e.g., electronic money), or something whose value varies only depending on its presence or absence (e.g., privileges). In the former case, the reward A to be given to the selected Y model learning devices 11-i is j(r+1) For example, the reward A is given in descending order of contribution to the associative learning. j (r+1) The size of may be determined.
[0045] <Example 2 of Reward Determination Method> The reward determination unit 127 determines the evaluation value e j (r+1) Reward A according to j (r+1) For example, the reward determination unit 127 may determine the magnitude of the evaluation value e j (r+1) Based on this, reward A is calculated according to the degree of contribution to the federated learning. j (r+1) For example, the magnitude of the evaluation value e j (r+1) indicates a greater contribution to associative learning, the reward determination unit 127 determines the evaluation value e j (r+1) The larger the reward A j (r+1) The magnitude of the evaluation value e j (r+1) A monotonically non-decreasing function value (a monotonically non-decreasing function value, for example, a monotonically increasing function value) is used as the reward A j (r+1) For example, the evaluation value e j (r+1) In this case, the smaller the evaluation value e, the greater the contribution to the associative learning. j (r+1) The smaller the reward A j (r+1) The magnitude of the evaluation value e j (r+1) A monotonically non-increasing function value (a monotonically non-increasing function value, for example, a monotonically decreasing function value) is used as the reward A. j (r+1) The size may be set to .
[0046] <Example 3 of Reward Determination Method> Example 1 and Example 2 of Reward Determination Method may be combined. For example, the reward determination unit 127 may j (r+1), 11-n, a model learning device 11-i to which a reward is to be given is selected from the plurality of model learning devices 11-1, . . . , 11-n, and a reward A to be given to the selected model learning device 11-i is calculated. j (r+1) The size of the evaluation value e j (r+1) Specific examples of the process are as exemplified in the examples 1 and 2 of the remuneration determination method.
[0047] The reward determination unit 127 determines reward A j (r+1) The communication unit 124 then transmits information representing the reward A j (r+1) The model learning device 11-j transmits information representing the reward A. j (r+1) The information representing the difference is received by the communication unit 114-i of the model learning device 11-j (FIG. 2). Note that the process of providing a reward will be omitted.
[0048] <Features of this embodiment> In this embodiment, the evaluation value e j (r+1) Therefore, it is possible to evaluate the contribution of model learning device 11-j to associative learning without incurring large computational costs.
[0049] In addition, since a large computational cost is not required to evaluate the contribution to federated learning, the evaluation value e j (r+1) That is, each time a learning round process including a learning process, an aggregation process, and an update process is performed, the evaluation value calculation unit 126 calculates the control coefficient c obtained in the learning round process. j (r+1) Based on this, the evaluation value e corresponding to the model learning device 11-j is calculated. j (r+1) can be obtained and output.
[0050] [First Modification of the First Embodiment] In the first embodiment, the evaluation value calculation unit 126 of the aggregation device 12 (FIG. 3) calculates the updated control coefficient c j (r+1) (new control coefficient) to calculate the evaluation value e j (r+1)and obtain the control coefficient c j (r+1) and ~ c (r+1) (new control coefficient) to calculate the evaluation value e j (r+1) However, the evaluation value calculation unit 126 calculates the control coefficient c before updating. j (r) Using the evaluation value e j (r+1) and obtain the control coefficient c j (r) and ~ c (r) Using the evaluation value e j (r+1) Alternatively, the evaluation value calculation unit 126 may obtain the control coefficient c j (r+1) and ~ c (r) Using the evaluation value e j (r+1) or the control coefficient c j (r) and ~ c (r+1) Using the evaluation value e j (r+1) That is, the evaluation value calculation unit 126 may obtain c j (r+1) Instead of c j (r) Using the evaluation value e j (r+1) You may also obtain ~ c (r+1) Instead of ~ c (r) Using the evaluation value e j (r+1) These processes may be performed by j (r+1) is c j (r) is replaced by ~ c (r+1) but ~ c (r) The same as that exemplified in the first embodiment may be used except that
[0051] [Modification 2 of the First Embodiment] Since a large computational cost is not required to evaluate the contribution to the federated learning, the evaluation value ej (r+1) and obtain the evaluation value e j (r+1) Remuneration A based on j (r+1) Therefore, the model learning device 11-i can determine the reward A j (r+1) In this example, the model learning device 11-i illustrated in FIG. 2 further includes a determination unit 117-i. In step S1226, the reward A transmitted from the aggregation device 12 is j (r+1) The information representing the reward A is received by the communication unit 114-i and sent to the determination unit 117-i. j (r+1) Whether to participate in the next learning round r+1 is determined based on the information representing the reward A, which has different values depending on its size. j (r+1) (for example, electronic money), the determination unit 117-i determines the reward A j (r+1) Compare the size of the reward A with a threshold. j (r+1) If the magnitude or average of the threshold value exceeds a threshold, it may be determined that the participant will participate in the next learning round r+1, and if not, it may be determined that the participant will not participate in the next learning round r+1. For example, the reward A j (r+1) (for example, a privilege), the determination unit 117-i determines a reward A j (r+1) If the reward A is given, the trainee may decide to participate in the next learning round r+1, and if not, the trainee may decide not to participate in the next learning round r+1. j (r+1) The information representing the reward A may be output from a user interface (display or speaker) not shown and presented to the user. In this case, the user may j (r+1)Depending on the instruction, model learning device 11-i may determine whether to participate in the next learning round r+1, and input information representing that instruction to determination unit 117-i. Determination unit 117-i may determine to participate in the next learning round r+1 if an instruction to participate in the next learning round r+1 has been input, and may determine not to participate in the next learning round r+1 if an instruction to participate in the next learning round r+1 has not been input. If determination unit 117-i determines to participate in the next learning round r+1, model learning device 11-i executes the processing of step 113-i in learning round r+1, and if not, model learning device 11-i does not execute the processing of step 113-i in learning round r+1.
[0052] [Third Modification of the First Embodiment] The model learning device 11-i receives the reward A transmitted from the aggregation device 12. j (r+1) Based on the information representing the local training data ξ to be used in step S1132a-i of the next training round r+1, i (r+1),k In this example, for example, a set of multiple types of local learning data having different characteristics is stored in the storage unit 113-i of the model learning device 11-i (FIG. 2). The set of multiple types of local learning data having different characteristics can be generated using techniques such as data cleaning, data augmentation, data normalization, feature engineering, and feature selection. In this example, the model learning device 11-i illustrated in FIG. 2 further includes a determination unit 117-i. In step S1226, the reward A transmitted from the aggregation device 12 is j (r+1) The information representing the reward A is received by the communication unit 114-i and sent to the determination unit 117-i. j (r+1) , from the sets of multiple types of local learning data stored in the storage unit 113-i, local learning data ξ to be used in step S1132a-i of the next learning round r+1 is selected based on the information representing i (r+1),k For example, the reward A has different values depending on its size. j (r+1) (for example, electronic money), the determination unit 117-i determines the reward Aj (r+1) Compare the size of the reward A with a threshold. j (r+1) If the magnitude or average of ξ exceeds a threshold, the learning data ξ used in step S1132a-i of the previous learning round r is i (r),k From the set of local training data of the same type as i (r+1),k On the other hand, for example, reward A j (r+1) If the magnitude or average of the learning data ξ used in step S1132a-i of the previous learning round r does not exceed the threshold, the determining unit 117-i i (r),k and the local training data set of different types, the local training data ξ to be used in step S1132a-i of the next training round r+1. i (r+1),k Select. Other rewards: j (r+1) The information representing the reward A may be output from a user interface (display or speaker) not shown and presented to the user. In this case, the user may j (r+1) The local training data ξ to be used in step S1132a-i of the next training round r+1 is i (r+1),k and inputs information representing the instruction to the determining unit 117-i. In accordance with the input instruction, the determining unit 117-i selects the local learning data ξ to be used in step S1132a-i of the next learning round r+1. i (r+1),k The learning unit 115-i selects the local learning data ξ selected by the determining unit 117-i. i (r+1),k Then, the process of steps S1132a-i of the next learning round r+1 is executed using the above.
[0053] Second Embodiment Next, a second embodiment will be described. In this embodiment, an evaluation method for performing associative learning different from that of the first embodiment will be illustrated. However, this does not limit the present invention. The following description will focus on differences from the first embodiment, and the same reference numbers will be used to simplify the description of matters already described.
[0054] <Configuration> As shown in FIG. 1, the federated learning system 2 of this embodiment has model learning devices 21-1, . . . , 21-n and an aggregation device 22 (evaluation device), which are configured to be able to communicate with each other via a network.
[0055] 6, the model learning device 21-i of this embodiment includes a memory 111-i, a control unit 112-i, a storage unit 113-i, a communication unit 114-i, a learning unit 215-i, and an update unit 216-i. The storage unit 113-i stores local learning data ξ i (r),k is stored.
[0056] 7 , the aggregating device 22 of this embodiment has a memory 121, a control unit 122, a communication unit 124, an aggregating unit 225, an evaluation value calculation unit 226, and a reward determination unit 127. The aggregating device 22 executes each process under the control of the control unit 122. Information input to the aggregating device 22 and information obtained by each unit is stored in the memory 121 one by one, and is read out as needed and used in each process.
[0057] <Processing> Next, an example of the federated learning process of this embodiment will be described. <Processing of Model Learning Device 21-i (FIG. 8)> Step S211-i: As shown in FIG. 8, the communication unit 114-i of the model learning device 21-i (where i∈{1, . . . , n}) (FIG. 6) participating in the federated learning receives the aggregated model parameters σ 1 , σ 2 , σ 3 , σ 4 , σ 5 , σ 6 , σ 6 , σ 7 , σ 8 , σ 9 , σ 10 , σ 11 , σ 2 , σ 3 , σ 4 , σ 5 , σ 6 , σ 10 , σ 2 , σ 3 , σ 4 , σ 5 , σ 6 , σ 10 , σ 2 , σ 3 , σ 2 , σ 3 , σ 4 , σ 2 , σ 3 ... ~ x (r) Initial value of ~ x (0) (r=0) is received. Initial value ~ x (0) is sent to the learning unit 215-i.
[0058] Step S112-i: The learning unit 215-i calculates the initial value of the aggregate model parameter. ~ x (0) The local model parameters x of the model learning device 11-i i (r) The initial value x i (0) (r=0). That is, the learning unit 215-i sets x i (0) = ~ x (0) Set to.
[0059] Step S213-i: The learning unit 215-i executes the following loop processing of steps S2131-i, S2132-i, S1133-i, S2134-i, S2135-i, and S2136-i for r = 0, ..., R-1. The learning unit 215-i first executes the loop processing for learning round r = 0, and then repeats the process of executing the loop processing for learning round r with r + 1 as the new r each time the loop processing for learning round r is executed, until the loop processing for learning round r = R-1 is completed.
[0060] Step S2131-i: The learning unit 215-i calculates the local model parameter x i (r) the local model parameters x of the local update step k of the learning round r i (r),k The initial value x i (r),0 (k=0). The learning unit 215-i sets the aggregate model parameter ~ x (r) is the local model parameter of local update step k in training round r. ~ x (r),k Initial value of ~ x (r),0 (k=0). That is, the learning unit 215-i sets x i (r),0 = x i (r) and ~ x (r),0 = ~ x (r) Set to.
[0061] Step S2132-i: The learning unit 215-i executes the loop processing of the following steps S2132a-i, S2132b-i, and S2132c-i for k = 0, ..., K-1. The learning unit 215-i first executes the loop processing of local update step k = 0, and then repeats the process of executing the loop processing of local update step k with k + 1 as the new k each time the loop processing of local update step k is executed, until the loop processing of local update step k = K-1 is completed.
[0062] Step S2132a-i: The learning unit 215-i calculates the local model parameter x i (r),k , the control coefficient c of the model learning device 21-i i (r) , the local learning data ξ read from the storage unit 113-i i (r),k Using the following calculation, the update amount v i (r),k+1 get. Here, the control coefficient c i (r) The initial value of c i (0) is a constant (e.g., a real number).
[0063] Step S2132b-i: The learning unit 215-i calculates the local model parameter x i (r),k , ~ x (r),k , update amount v i (r),k+1 , and a constant η, the following calculation is performed to obtain the local model parameters x i (r),k+1 get. Here, the constant η is a real number greater than 0 and less than 1.
[0064] Step S2132c-i: The learning unit 215-i i (r),k+1 of ~ x (r),k+1 That is, the learning unit 215-i sets the following:
[0065] Step S1133-i: The learning unit 215-i i (r),K the new local model parameters x i (r+1) That is, the learning unit 215-i sets x i (r+1) = x i (r),K In this way, the learning unit 215-i sets at least the aggregate model parameter ~ x (r) , the local learning data ξ of the model learning device 21-i i (r),k , and the control coefficient c of the model learning device 21-i i (r) Based on the aggregate model parameters ~ x (r) The local model parameters x of the model learning device 21-i i (r+1) The learning unit 215-i executes a learning process to update the new local model parameter x i (r+1) to the update unit 216-i and the communication unit 114-i. i (r) to the update unit 216-i.
[0066] Step S2134-i: The communication unit 114-i receives the new local model parameter x i (r+1) is transmitted to the aggregation device 22 via the network.
[0067] Step S1135-i The communication unit 114-i receives the new aggregation model parameters transmitted from the aggregation device 22 via the network. ~ x (r+1) Receive the aggregate model parameters ~ x (r+1) are sent to the learning unit 215-i and the updating unit 216-i. ~ x (r+1) The generation of will be described later.
[0068] Step S2136-i The update unit 216-i updates the aggregate model parameters~ x (r+1) , new local model parameters x i (r+1) , and the control coefficient c i (r) Using the following calculation, a new control coefficient c i (r+1) and output it. Here, K η is a positive real constant. Also, the control coefficient c i (r) The initial value of c i (0) (r=0) is a constant (for example, a real number) greater than or equal to 0. That is, the update unit 216-i updates at least the new aggregate model parameters ~ x (r+1) and the new local model parameters x i (r+1) Based on the difference between i (r) The new control coefficient c i (r+1) The update unit 216-i executes an update process to update the new control coefficient c i (r+1) is sent to the learning unit 215-i.
[0069] <Processing of Aggregation Device 22 (FIG. 9)> Step S221: As illustrated in FIG. 9, the communication unit 124 of the aggregation device 22 (FIG. 7) receives the above-mentioned aggregation model parameters ~ x (r) Initial value of ~ x (0) (r=0) is transmitted via the network to model learning device 21-j participating in federated learning, where j∈N and N⊆{1,...,n}.
[0070] Step S222: The aggregation device 22 executes the loop processing of the following steps S2221, S1222, S2224, S2225, and S1226 for r = 0, ..., R-1. The aggregation device 22 first executes the loop processing of learning round r = 0, and repeats the process of executing the loop processing of learning round r with r + 1 as a new r each time the loop processing of learning round r is executed, until the loop processing of learning round r = R-1 is completed.
[0071] Step S2221: The communication unit 124 of the aggregation device 22 (FIG. 7) receives the local model parameters x transmitted as described above via the network from the model learning device 21-j participating in the federated learning. j (r+1) Receive the local model parameters x j (r+1) is sent to the aggregation unit 225.
[0072] Step S1222: The aggregating unit 225 aggregates the sent local model parameters x j (r+1) (where j∈N), and calculate the aggregate model parameters in training round r+1 by ~ x (r+1) get. Here, |N| represents the number of elements of N, and represents the number of model learning devices 21-j participating in the federated learning. That is, the aggregator 225 aggregates the local model parameters x j (r+1) Aggregated model parameters using ~ x (r) the new aggregated model parameters ~ x (r+1) Aggregation model parameters are updated to ~ x (r+1) is sent to the communication unit 125. In addition, the local model parameter x j (r+1) and aggregate model parameters ~ x (r+1) is sent to the evaluation value calculation unit 226.
[0073] Step S2224: The communication unit 125 transmits the aggregate model parameters~ x (r+1) For example, the communication unit 125 transmits the ~ x (r+1) Alternatively, the model learning device 21-j may transmit the ~ x (r+1) may be transmitted.
[0074] Step S2225: The evaluation value calculation unit 226 calculates the evaluation value e corresponding to the model learning device 21-i based on the control coefficient. j (r+1) In this embodiment, for example, the evaluation value calculation unit 226 obtains and outputs the above-mentioned initial value c j (0) , real constant K η , local model parameters x j (r+1) , aggregate model parameters ~ x (r+1) Using the following calculation, a new control coefficient c j (r+1) and this control coefficient c j (r+1) Based on this, the evaluation value e corresponding to the model learning device 21-j is calculated. j (r+1) and output it. Alternatively, the evaluation value calculation unit 226 may further calculate the control coefficient ~ c (r+1) Obtain c j (r+1) and ~ c (r+1) Using the evaluation value e j (r+1) may be obtained. Evaluation value e j (r+1) A specific example of the evaluation value e is as exemplified in step S1225 of the first embodiment. j (r+1) is sent to the remuneration determination unit 127.
[0075] Step S1226: The reward determination unit 127 calculates the evaluation value e j (r+1) The reward A for the model learning device 11-j isj (r+1) The reward determination unit 127 determines and outputs the reward A (incentive). j (r+1) The communication unit 124 then transmits information representing the reward A j (r+1) The model learning device 21-j transmits information representing the reward A j (r+1) The information representing is received by the communication unit 114-i of the model learning device 21-j (FIG. 6).
[0076] <Features of this embodiment> In this embodiment, the evaluation value e j (r+1) Therefore, it is possible to evaluate the contribution of model learning device 11-j to associative learning without incurring large computational costs.
[0077] In addition, since a large computational cost is not required to evaluate the contribution to federated learning, the evaluation value e j (r+1) That is, each time a learning round process including a learning process, an aggregation process, and an update process is executed, the evaluation value calculation unit 226 calculates the control coefficient c obtained in the learning round process. j (r+1) Based on this, the evaluation value e corresponding to the model learning device 21-j is calculated. j (r+1) can be obtained and output.
[0078] [Modification 1 of the Second Embodiment] In step S2225 of this embodiment, the evaluation value calculation unit 226 calculates c j (r+1) However, the evaluation value calculation unit 226 calculated j (r+1) Instead of calculating j (r+1) is transmitted to the aggregation device 22, and the evaluation value calculation unit 226 calculates the transmitted c j (r+1) Using the evaluation value e j (r+1) may be obtained.
[0079] [Modification 2 of Second Embodiment] As in Modification 1 of the first embodiment, in the second embodiment, the evaluation value calculation unit 226 calculates the value of j (r+1) Instead of c j (r) Using the evaluation value e j (r+1) You may also obtain ~ c (r+1) Instead of ~ c (r) Using the evaluation value e j (r+1) may be obtained.
[0080] [Modification 3 of Second Embodiment] In the second embodiment, since a large computational cost is not required to evaluate the contribution to the federated learning, the evaluation value e j (r+1) and obtain the evaluation value e j (r+1) Remuneration A based on j (r+1) Therefore, similarly to the second modification of the first embodiment, the model learning device 21-i can determine the reward A j (r+1) In addition, similar to the third modification of the first embodiment, the model learning device 21-i can determine whether to participate in the next learning round r+1 based on the information representing the reward A j (r+1) Based on the information representing the local training data ξ to be used in step S2132a-i of the next training round r+1, i (r+1),k may be selected.
[0081] [Third Embodiment] In the first and second embodiments and their modified examples, aggregating device 12 includes evaluation value calculation unit 126 and reward determination unit 127, and aggregating device 22 includes evaluation value calculation unit 226 and reward determination unit 127. Evaluation value calculation unit 126 and reward determination unit 127 may be provided outside aggregating device 12, or evaluation value calculation unit 226 and reward determination unit 127 may be provided outside aggregating device 22. That is, as illustrated in FIG. 1 , federated learning system 3 may include evaluation device 33 including evaluation value calculation unit 126 and reward determination unit 127, model learning devices 11-1, ..., 11-N, and aggregating device 12, or federated learning system 3 may include evaluation device 33 including evaluation value calculation unit 226 and reward determination unit 127, model learning devices 21-1, ..., 21-N, and aggregating device 22. In this case, the evaluation value calculation units 126, 226 and the remuneration determination unit 127 may transmit and receive information via a network and execute the processes of the evaluation value calculation units 126, 226 and the remuneration determination unit 127. In this case, the evaluation value calculation unit 126 and the remuneration determination unit 127 may be omitted from the aggregation device 12, and the evaluation value calculation unit 226 and the remuneration determination unit 127 may be omitted from the aggregation device 22.
[0082] [Fourth Embodiment] In the first and second embodiments and their modified examples, the aggregation device 12 and the aggregation device 22 are provided external to the model learning devices 11-1, ..., 11-N and the model learning devices 21-1, ..., 21-N. However, any one of the model learning devices 11-1, ..., 11-N may have the functional configuration of the aggregation device 12, or any one of the model learning devices 21-1, ..., 21-N may have the functional configuration of the aggregation device 22. In such cases, the aggregation device 12 and the aggregation device 22 may be omitted. Furthermore, in such cases, the evaluation value calculation unit 126 and the reward determination unit 127 may be provided external to the model learning devices 11-1, ..., 11-N, or the evaluation value calculation unit 226 and the reward determination unit 127 may be provided external to the model learning devices 21-1, ..., 21-N.
[0083] [Hardware Configuration] The model learning devices 11-i and 21-i, the aggregation devices 12 and 22, and the evaluation device 33 in each embodiment are devices configured by a general-purpose or dedicated computer having, for example, a processor (hardware processor) such as a central processing unit (CPU) and memories such as random-access memory (RAM) and read-only memory (ROM) executing a predetermined program. That is, the model learning devices 11-i and 21-i, the aggregation devices 12 and 22, and the evaluation device 33 in each embodiment have, for example, processing circuitry configured to implement each of the components possessed by each of them. This computer may have one processor and memory, or multiple processors and memories. This program may be installed on the computer or may be pre-recorded in a ROM or the like. Furthermore, some or all of the processing units may be configured using electronic circuits that independently realize processing functions, rather than electronic circuits that realize functional configurations by loading programs like a CPU. Furthermore, the electronic circuits constituting a single device may include multiple CPUs.
[0084] FIG. 6 is a block diagram illustrating the hardware configuration of the model learning devices 11-i and 21-i, the aggregation devices 12 and 22, and the evaluation device 33 in each embodiment. As illustrated in FIG. 6, the model learning devices 11-i and 21-i, the aggregation devices 12 and 22, and the evaluation device 33 in this example include a central processing unit (CPU) 10a, an input unit 10b, an output unit 10c, a random access memory (RAM) 10d, a read-only memory (ROM) 10e, an auxiliary storage device 10f, a communication unit 10h, and a bus 10g. The CPU 10a in this example includes a control unit 10aa, a calculation unit 10ab, and a register 10ac, and performs various calculation processes according to various programs loaded into the register 10ac. The input unit 10b is an input terminal, keyboard, mouse, touch panel, etc., through which data is input. The output unit 10c is an output terminal, display, etc., through which data is output. The communication unit 10h is a LAN card or the like controlled by the CPU 10a that has loaded a predetermined program. The RAM 10d is a static random access memory (SRAM) or dynamic random access memory (DRAM), and has a program area 10da where the predetermined program is stored and a data area 10db where various data is stored. The auxiliary storage device 10f is a hard disk, magneto-optical disk (MO), semiconductor memory, or the like, and has a program area 10fa where the predetermined program is stored and a data area 10fb where various data is stored. The bus 10g connects the CPU 10a, input unit 10b, output unit 10c, RAM 10d, ROM 10e, communication unit 10h, and auxiliary storage device 10f so that information can be exchanged. The CPU 10a writes the program stored in the program area 10fa of the auxiliary storage device 10f to the program area 10da of RAM 10d in accordance with the loaded OS (Operating System) program. Similarly, the CPU 10a writes various data stored in the data area 10fb of the auxiliary storage device 10f to the data area 10db of the RAM 10d.The addresses on RAM 10d where these programs and data are written are then stored in register 10ac of CPU 10a. Control unit 10aa of CPU 10a sequentially reads these addresses stored in register 10ac, reads the programs and data from the areas on RAM 10d indicated by the read addresses, causes calculation unit 10ab to sequentially execute the calculations indicated by the programs, and stores the calculation results in register 10ac. This configuration realizes the functional configurations of model learning devices 11-i and 21-i, aggregation devices 12 and 22, and evaluation device 33.
[0085] The above-mentioned program can be recorded on a computer-readable recording medium. Examples of computer-readable recording media include non-transitory recording media. Examples of such recording media include magnetic recording devices, optical disks, magneto-optical recording media, and semiconductor memories.
[0086] This program may be distributed, for example, by selling, transferring, or lending a portable recording medium, such as a DVD or CD-ROM, on which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of a server computer and transferring the program from the server computer to other computers via a network. As described above, a computer that executes such a program may first temporarily store the program recorded on a portable recording medium or transferred from the server computer in its own storage device. Then, when executing a process, the computer reads the program stored in its own storage device and executes processing in accordance with the read program. Alternatively, the program may be executed by a computer that reads the program directly from a portable recording medium and executes processing in accordance with the program. Furthermore, the computer may execute processing in accordance with the received program each time a program is transferred from the server computer to the computer. Alternatively, the server computer may not transfer the program to the computer, but may instead execute the processing function simply by issuing an execution instruction and obtaining the results, thereby executing the processing described above through a so-called ASP (Application Service Provider) service. In this embodiment, the program includes information used for processing by an electronic computer that is equivalent to a program (such as data that is not a direct instruction to a computer but has properties that dictate computer processing).
[0087] In each embodiment, the device is configured by executing a predetermined program on a computer, but at least a part of the processing contents may be realized by hardware.
[0088] The present invention is not limited to the above-described embodiment. For example, an evaluation value is obtained each time the learning round process is executed β times, and a reward A based on the evaluation value is calculated. j (r+1)Here, β is an integer of 2 or more. Furthermore, the various processes described above may not only be executed in chronological order as described, but may also be executed in parallel or individually depending on the processing capacity of the device executing the processes or as necessary. Needless to say, other appropriate modifications are possible within the scope of the present invention.
[0089] 1-3 Federated learning system 11, 21 Model learning device 12, 22 Aggregation device (evaluation device) 33 Evaluation device 126, 226 Evaluation value calculation unit 127 Reward determination unit
Claims
1. An evaluation device having: an input unit that accepts a value corresponding to the control coefficient or a value corresponding to the new control coefficient obtained by associative learning that repeats a learning process that updates the aggregate model parameters to local model parameters of each of the model learning devices based on an aggregate model parameter, local learning data of each of a plurality of model learning devices, and a control coefficient of each of the model learning devices; an aggregation process that updates the aggregate model parameters to new aggregate model parameters using the local model parameters; and an update process that updates the control coefficients to new control coefficients based on a difference between the aggregate model parameters or the new aggregate model parameters and the local model parameters or the new local model parameters; and an evaluation value calculation unit that obtains and outputs an evaluation value corresponding to the model learning device based on the control coefficients or the new control coefficients.
2. An evaluation device as claimed in claim 1, wherein the evaluation value is a first evaluation value when the control coefficient or an index based on the new control coefficient is a first value, the evaluation value is a second evaluation value when the index is a second value, and the evaluation value is a third evaluation value when the index is a third value, the first value is smaller than the second value, the second value is smaller than the third value, the first evaluation value represents a greater contribution to the associative learning than the second evaluation value, and / or the third evaluation value represents a greater contribution to the associative learning than the second evaluation value.
3. An evaluation device according to claim 2, wherein the index is a value representing the magnitude of the control coefficient or the new control coefficient.
4. An evaluation device according to claim 2, wherein the index is a relative value of the amount of information of the control coefficient of each of the model learning devices with respect to a statistical value of the amount of information of the control coefficients of the plurality of model learning devices.
5. An evaluation device according to claim 1, wherein the evaluation value calculation unit obtains and outputs an evaluation value corresponding to the model learning device based on the control coefficient or the new control coefficient obtained in the learning round process each time a learning round process including the learning process, the aggregation process, and the update process is executed once.
6. The evaluation device of claim 1, further comprising a reward determination unit that (1) selects a model learning device to which a reward will be given from among the plurality of model learning devices based on the evaluation value, and / or (2) determines the magnitude of the reward in accordance with the evaluation value.
7. An evaluation method using an evaluation device, comprising: an input step in which a value corresponding to the control coefficient or a value corresponding to the new control coefficient obtained by repeated associative learning that includes a learning process for updating the aggregate model parameters to local model parameters of each of the model learning devices based on aggregate model parameters, local learning data of each of a plurality of model learning devices, and a control coefficient of each of the model learning devices, an aggregation process for updating the aggregate model parameters to new aggregate model parameters using the local model parameters, and an update process for updating the control coefficients to new control coefficients based on the difference between the aggregate model parameters or the new aggregate model parameters, and the local model parameters or the new local model parameters, is input to an input unit; and an evaluation value calculation step in which an evaluation value calculation unit obtains and outputs an evaluation value corresponding to the model learning device based on the control coefficients or the new control coefficients.
8. A program for causing a computer to function as the evaluation device according to any one of claims 1 to 6.
Citation Information
Patent Citations
Medical information processing apparatus, medical information processing system, and medical information processing method
JP2021056995A
Evaluating a contribution of participants in federated learning
US20220383202A1
