Method for calculating contribution in associative learning and a system for calculating contribution and profit sharing in associative learning

Abstract

To provide a method and system for calculating contributions in federated learning and a system for performing profit-sharing on the basis of the calculated contributions.SOLUTION: A method comprises the steps of: collaboratively developing, by all of participants in federated learning, a federated aggregation model by federated learning, using their own local datasets; excluding participation of at least one first member in all participants, and then the remained participants excluding the at least one first member collaboratively developing a first contribution model by a federated learning method; and, comparing the value of the first contribution model with the value of the federated aggregation model to obtain the contribution of the first member. The method of the present invention is capable of calculating the contribution of each of the members or group contributions of some members in the federated learning by few additional information and few additional calculations, so as to achieve profit sharing according to the contributions.SELECTED DRAWING: Figure 2

Description

Technical Field

[0001] The present invention relates to a method for calculating contribution degrees in federated learning and a contribution degree calculation and profit distribution system in federated learning, and particularly to a method and system capable of calculating the contribution degrees of each member in federated learning, and a system capable of performing profit distribution based on the calculated contribution degrees.

Background Art

[0002] With the progress of medical engineering technology, currently, the application of regenerative medicine in the clinical treatment of diseases is becoming more diverse, and its research and application are attracting increasing attention from all sectors. Currently, regenerative medicine is mainly applied to organ repair, immune cell therapy, and stem cell therapy. Cell therapy refers to a process in which human cells are cultured or processed outside the body and then transplanted into individuals' bodies for use. Therefore, how to effectively culture cells is an extremely important factor in regenerative medicine. From previous research, it has been found that cells in each person's body have unique characteristics, and it is necessary to customize cells suitable for treatment according to each situation, so the complexity and difficulty of the cell process increase. On the other hand, when there is a large amount of cell data and bioprocess data, it is possible to generate a prediction model for evaluating the effects of the conditions and process parameters of each cell culture by machine learning. Therefore, by using machine learning, the complexity and difficulty in customizing the cell process can be reduced. However, most of these cell data and bioprocess data belong to very delicate and highly confidential personal data, such as the cell data of patients in various hospitals or laboratories. When using the conventional centralized machine learning method, it is necessary to aggregate all bioprocess data on a central server for analysis and training, so it may be impossible to implement due to restrictions such as regulations, privacy protection, and commercial considerations. Also, not only in medical engineering, but in any technical field where data is very delicate and confidential, the centralized machine learning method may face the same problems. [

[0003] Federated learning, also known as collaborative learning, is a type of machine learning technique that specifically involves training an algorithm using multiple distributed edge devices or servers that possess local data samples. This method differs significantly from traditional centralized machine learning techniques. In traditional centralized machine learning, all local datasets are uploaded to a single server for machine learning training. In contrast, federated learning consists of multiple members. Each member trains its own dataset to generate its own trained model and uploads the trained model parameters to a central server. The central server then assigns weights to all the trained model parameters, aggregates them, and forms a federated model. This federated model can then be redistributed to each member for use. The central server only aggregates the trained model parameters and does not directly perform any machine learning analysis or processing on each member's dataset. Therefore, highly sensitive and confidential data samples and information within each member's dataset are not leaked externally, and data confidentiality is protected.

[0004] The associative model ultimately generated through associative learning is completed by the models that each member trains and contributes. Therefore, each member has its own contribution to the associative model. Furthermore, if the associative model is commercially used due to its reliability and predictive accuracy and is profitable, each member may seek to receive a share of the profits according to their respective contribution. In addition, each member's contribution also represents their respective influence on the associative model, and can be used to evaluate whether the weights of each member in the associative model are appropriate. However, currently, there is no method in associative learning that allows for the calculation of each member's contribution.

[0005] Therefore, in order to apply this to profit distribution to members and other functions based on contribution, it is necessary to develop a method and system that can efficiently calculate the contribution of each member in associative learning. [Overview of the Initiative] [Problems that the invention aims to solve]

[0006] In view of the above, the present invention provides a method for calculating the contribution in associative learning in order to solve the conventional problems described above. [Means for solving the problem]

[0007] According to a specific embodiment of the present invention, a method for calculating contribution in federated learning includes the steps of: a central server receiving a plurality of trained model parameters, the trained model parameters being generated by machine learning on a plurality of local datasets; the central server generating a federated model by executing an aggregation algorithm on the trained model parameters and saving the federated model; the central server generating a first contribution model by executing an aggregation algorithm on the trained model parameters excluding or simulating at least one first trained model parameter; and the central server generating the contribution of the first trained model parameter by comparing the value of the first contribution model with the value of the federated model.

[0008] The aforementioned trained model parameters are generated by performing multiple machine learning training sessions on the local dataset. The central server generates a federated model by executing an aggregation algorithm in each machine learning training session. Furthermore, the central server generates a first contribution model by executing an aggregation algorithm on the trained model parameters, excluding at least one first trained model parameter from the last machine learning training session.

[0009] The aforementioned trained model parameters are generated by performing multiple machine learning training sessions on the local dataset. The central server generates a federated model by executing an aggregation algorithm in each machine learning training session. Furthermore, the central server generates a first contribution model by executing an aggregation algorithm on the trained model parameters, excluding at least one first trained model parameter from each machine learning training session in a simulated manner.

[0010] The trained model parameters are generated by performing multiple machine learning training sessions on the local dataset. The central server generates a federated model by executing the aggregation algorithm in each machine learning training session. The central server also generates a first contribution model by executing the aggregation algorithm on the trained model parameters, excluding at least one first trained model parameter from the last k machine learning training sessions (counting down from the last session). k is an integer between 1 and the number of times the first trained model parameter participated in the aggregation algorithm of the federated model.

[0011] The aforementioned trained model parameters are generated by performing multiple machine learning training sessions on the local dataset. The central server generates a federated model corresponding to each machine learning session by executing an aggregation algorithm in each machine learning training session. Furthermore, it generates a first contribution model corresponding to each machine learning session by executing an aggregation algorithm on the trained model parameters, excluding at least one first trained model parameter from each machine learning training session. In addition, the value of the first contribution model in each machine learning session is compared with the value of the federated model, and the average or weighted average is calculated to generate the average contribution as the contribution of the first trained model parameter.

[0012] The method further includes the steps of: a central server determining whether the contribution of at least one first trained model parameter is positive or negative; and, if the contribution is negative, the central server replacing the federated model with the first contribution model to obtain the updated federated model.

[0013] Furthermore, the present invention solves the aforementioned conventional problems and provides a contribution calculation and benefit distribution system for associative learning that can be applied to benefit distribution among members in associative learning.

[0014] According to a specific embodiment of the present invention, the contribution calculation and profit distribution system in federated learning includes a plurality of artificial intelligence model training devices and a central server. The central server further includes a computing module, a contribution calculation module, and a profit distribution module. The plurality of artificial intelligence model training devices generate a plurality of trained model parameters by machine learning a plurality of local datasets. The central server is connected to each artificial intelligence model training device and receives the trained model parameters. The computing module of the central server generates a federated model by executing an aggregation algorithm on the received trained model parameters. The contribution calculation module generates a first contribution model by executing an aggregation algorithm on the trained model parameters, excluding or simulating at least one first trained model parameter. Then, it compares the value of the first contribution model with the value of the federated model to generate the contribution of the first trained model parameter. The profit distribution module is connected to the contribution calculation module and receives the contribution of the first trained model parameter, and calculates the profit distribution ratio of the first trained model parameter in the federated model based on the contribution.

[0015] The artificial intelligence model training device performs multiple machine learning training sessions on the local dataset. The computing module generates a federated model by executing an aggregation algorithm in each machine learning training session. The contribution calculation module generates a first contribution model by executing an aggregation algorithm on the trained model parameters, excluding at least one first trained model parameter from the last machine learning training session.

[0016] The artificial intelligence model training device performs multiple machine learning training sessions on the local dataset. The computing module generates a federated model by executing an aggregation algorithm in each machine learning training session. The contribution calculation module generates a first contribution model by executing an aggregation algorithm on the trained model parameters, excluding at least one first trained model parameter from each machine learning training session in a simulated manner.

[0017] The artificial intelligence model training device performs multiple machine learning training sessions on the local dataset. The computing module generates the federated model by executing an aggregation algorithm in each machine learning training session. The contribution calculation module generates a first contribution model by executing an aggregation algorithm on the trained model parameters, which are obtained by simulating the exclusion of at least one first trained model parameter from the last k machine learning training sessions, counting down from the last session. k is an integer between 1 and the number of times the first trained model parameter participated in the aggregation algorithm of the federated model.

[0018] The artificial intelligence model training device performs multiple machine learning training sessions on the local dataset. The computation module generates a federated model corresponding to each machine learning session by executing the aggregation algorithm in each machine learning training session. The contribution calculation module generates a first contribution model corresponding to each machine learning session by executing the aggregation algorithm on the trained model parameters, excluding at least one first trained model parameter from each machine learning training session in a simulated manner. Furthermore, the value of the first contribution model in each machine learning session is compared with the value of the federated model, and the average or weighted average is calculated to generate the average contribution as the contribution of the first trained model parameter. [Effects of the Invention]

[0019] In summary, the present invention provides a method and system for calculating contributions in associative learning that can calculate the contribution of each member participating in associative learning to the associated model. This also enables the function of distributing benefits among the members. Alternatively, by adjusting the weights of each member's model in the associated model based on their contribution, it is possible to give the associated model generated by associative learning a more accurate predictive effect. [Brief explanation of the drawing]

[0020] [Figure 1] Figure 1 shows a flowchart of the steps for calculating the contribution in associative learning based on a specific embodiment of the present invention. [Figure 2] Figure 2 shows a schematic diagram of a contribution calculation and profit distribution system in federative learning based on another specific embodiment of the present invention. [Figure 3] Figure 3 shows a flowchart of the steps for calculating the contribution in associative learning based on another specific embodiment of the present invention. [Figure 4] Figure 4 shows a flowchart of the steps for calculating the contribution in associative learning based on another specific embodiment of the present invention. [Figure 5] FIG. 5 shows a flowchart of steps of a method for calculating contribution degrees in collaborative learning based on another specific embodiment of the present invention. [Figure 6] FIG. 6 shows a flowchart of steps of a method for calculating contribution degrees in collaborative learning based on another specific embodiment of the present invention. [Figure 7] FIG. 7 shows a flowchart of steps of a method for calculating contribution degrees in collaborative learning based on another specific embodiment of the present invention. [Figure 8] FIG. 8 shows a flowchart of steps of a method for calculating contribution degrees in collaborative learning based on another specific embodiment of the present invention. **DETAILED DESCRIPTION OF THE INVENTION**

[0021] Subsequently, specific embodiments will be used and the drawings will be referred to for a detailed description and discussion so that the advantages, spirit, and features of the present invention can be more easily and clearly understood. It should be noted that these specific embodiments are merely representative specific embodiments of the present invention, and the specific methods, apparatuses, conditions, materials, etc. exemplified do not limit the present invention or the corresponding specific embodiments. Also, each component in the drawings is used only to represent their relative positions and is not described based on actual ratios. Further, the step numbers of the present invention are only for distinguishing different steps and do not represent the order of the steps. This is explained in advance above.

[0022] Refer to FIGS. 1 and 2. FIG. 1 shows a flowchart of steps of a method for calculating contribution degrees in collaborative learning based on a specific embodiment of the present invention. FIG. 2 shows a schematic diagram of a contribution degree calculation / benefit distribution system 2 in collaborative learning based on another specific embodiment of the present invention. It should be noted that each step of the method in FIG. 1 can be executed by different units of the system in FIG. 2. Therefore, hereinafter, each step of the method in FIG. 1 will be described using the system in FIG. 2.

[0023] As shown in Figure 2, the contribution calculation and profit distribution system 2 in federated learning of this specific embodiment includes a plurality of artificial intelligence model training devices 20 and a central server 22 connected to these artificial intelligence model training devices 20. Each artificial intelligence model training device 20 performs machine learning on its respective local dataset to generate its own trained model. In practice, the artificial intelligence model training devices 20 are distributed nodes capable of independently executing complete machine learning algorithms. For example, in the field of regenerative medicine described above, they can be computer equipment capable of executing machine learning algorithms in each hospital or laboratory. The local dataset refers to a sample dataset for training, verification, and testing held by each hospital or laboratory. Each artificial intelligence model training device 20 can transmit the model parameters it has trained to the central server. Although only two artificial intelligence model training devices 22 are shown in Figure 2, in practice, the invention is not limited to these, and any number of artificial intelligence model training devices and the trained model parameters generated by these devices can be used in this invention. The central server 22 is capable of accepting trained model parameters trained by these artificial intelligence model training devices 20, and includes a computing module 220, a contribution calculation module 222, and a profit distribution module 224. The contribution calculation module 222 is connected to the computing module 220, and the profit distribution module 224 is connected to the contribution calculation module 222. In practice, the computing module 220, the contribution calculation module 222, and the profit distribution module 224 are built to fit within the computing unit of the central server 22, such as a computer device, or the central processing unit of a server, or a computing chip. Alternatively, the functions of the computing module 220, the contribution calculation module 222, and the profit distribution module 224 may be performed by different computing units connected to each other.

[0024] The computation module 220 can generate an associative model by executing an aggregation algorithm on all accepted trained model parameters. In the actual operation of the aggregation algorithm, the computation module 220 assigns weights to each trained model parameter before aggregation is possible. Note that the aggregation algorithm for associative learning is common knowledge in this field and will not be described in further detail here. It should be noted that any aggregation algorithm for generating an associative model is applicable to the present invention. The contribution calculation module 222 first simulates excluding at least one first trained model parameter from all trained model parameters, and then obtains a first contribution model by executing the same aggregation algorithm on the remaining trained model parameters after excluding or simulcasting the at least one first trained model parameter. Subsequently, the contribution of the first trained model parameter to the associative model can be determined by comparing the value of the first contribution model with the value of the associative model. In detail, if the first trained model parameters contribute to the federated model, then if the first trained model parameters are excluded from all trained model parameters, the value of the first contribution model generated by a similar aggregation algorithm should be lower than that of the federated model. Therefore, it is possible to estimate the contribution of the first trained model parameters from the reduced value. The values ​​of the first contribution model and the federated model can be estimated from the accuracy of each model. For example, if a value function is defined in which the relationship between value and the accuracy of the model is linear, it will be shown that the higher the accuracy of the model, the higher the value. In practice, the value function can be jointly determined by all federated members and is not limited to the linear relationship described above.

[0025] The profit-sharing module 224 can determine the profit-sharing ratio of the artificial intelligence model training device that provided the first trained model, based on the contribution or collective contribution of at least one first trained model parameter. Therefore, if the federated model generates profits or revenues, the members of the artificial intelligence model training device that generated the first trained model parameters can receive profits based on the profit-sharing ratio. In practice, the profit-sharing module 224 uses an algorithm for calculating the profit-sharing ratio based on the Shapley value or marginal contribution, which are criteria for fairness, to determine the profit-sharing ratio of the first trained model parameter or all trained model parameters based on their contributions.

[0026] As described above, the contribution calculation module 222 excludes or simulates at least one first-trained model parameter. If at least one first-trained model parameter is a set of first-trained model parameters, that is, if the participation of one federation member is simulated, the individual contribution of the excluded federation member is calculated after comparing the value of the generated first-contribution model with that of the federation model. In contrast, if at least one first-trained model parameter is a set of multiple first-trained model parameters, that is, if the participation of multiple federation members is excluded or simulated simultaneously, the collective contribution of the excluded federation members is calculated after comparing the value of the generated first-contribution model with that of the federation model. It should be noted that the collective contribution calculated by simultaneously excluding multiple federation members is not necessarily equal in practice to the sum of the individual contributions calculated when those federation members are individually simulated and excluded.

[0027] The method for calculating the contribution of the contribution calculation and profit distribution system 2 in the aforementioned associative learning is shown in Figure 1. In the specific embodiment shown in Figure 1, the method for calculating the contribution in associative learning includes the following steps.

[0028] Step S10: The central server 22 accepts multiple trained model parameters. These trained model parameters are generated by machine learning on multiple local datasets.

[0029] Step S12: The central server 22 generates and saves a federated model by running an aggregation algorithm on all trained model parameters.

[0030] Step S14: The central server 22 generates a first contribution model by running an aggregation algorithm on the other trained model parameters, excluding or simulating at least one first trained model parameter.

[0031] Step S16: The central server 22 compares the value of the first contribution model with the value of the federated model to generate the contribution of at least one first trained model parameter (i.e., generates the contribution of at least one federated member of the first trained model parameter).

[0032] In step S10, the central server 22 can accept trained model parameters from multiple artificial intelligence model training devices 20 as shown in Figure 2. In step S12, the central server 22 can generate a federated model by assigning weights to each trained model parameter using an aggregation algorithm and aggregating them using the computing module 220. In step S14, the contribution calculation module 222 of the central server 22 can first simulate the exclusion of at least one first trained model parameter from all the trained model parameters. Next, it obtains a first contribution model by running the aggregation algorithm on the remaining trained model parameters from which at least one first trained model parameter has been simulated excluded. In step S16, the contribution calculation module 222 compares the value of the first contribution model with the value of the federated model and can inversely estimate the individual or collective contribution of at least one first trained model parameter based on the degree of value decrease.

[0033] It should be noted that, in the specific embodiment described above, the trained model parameters provided by each member in federated learning are aggregated by a central processing unit. However, the central processing unit is merely a representative computing device for aggregating and generating the federated model and the first contribution model in this invention, and is not limited to a central processing unit in the narrow sense. In other words, in practical use, the central processing unit may be a cloud server. In this case, the trained model parameters of the federated members are received via the internet, aggregated into the federated model and the first contribution model, and then the federated model is redistributed to each federated member via the internet. Then, the contribution of each federated member is obtained by performing the calculations described above.

[0034] The contribution calculation and profit sharing system 2 in federative learning shown in Figure 2 can calculate the individual or collective contribution of the trained models generated by each member or group of members in federative learning, using the contribution calculation method in federative learning shown in Figure 1. Based on these contributions, the profit sharing ratio of each member or group of members can be determined. However, the calculated contribution of each trained model is not only usable for the profit sharing function. Now, refer to Figure 3. Figure 3 shows a flowchart of the steps of the contribution calculation method in federative learning based on another specific embodiment of the present invention. Each step of the method in Figure 3 can also be performed by different units of the system in Figure 2. Therefore, below, each step of the method in Figure 3 will be described using the system in Figure 2. As shown in Figure 3, this specific embodiment differs from the specific embodiment described above in the following respects. That is, the method of this specific embodiment further includes the following steps.

[0035] Step S17: The central server 22 adjusts the weights of the first trained model parameters in the federated model based on the contribution of the first trained model parameters.

[0036] The contribution is determined by the value of the trained model parameters to the associative model. Parameters with a high contribution have a large impact on the value of the associative model, while those with a low contribution have a small impact. Therefore, the associative model can be made more accurate by readjusting the weights based on the contribution of each trained model parameter. It should be noted that the other steps in this specific example are almost the same as the corresponding steps in the specific example described above, and will not be described in detail again here.

[0037] Furthermore, if a trained model is found to have a negative contribution, it means that the trained model has a negative impact on the associative model and should therefore be excluded. Refer to Figure 4. Figure 4 shows a flowchart of the steps for calculating the contribution in associative learning based on another specific embodiment of the present invention. Each step of the method in Figure 4 can be performed by different units of the system in Figure 2. Therefore, each step of the method in Figure 4 will be described below using the system in Figure 2. As shown in Figure 4, the method of this specific embodiment further includes the following steps.

[0038] Step S180: The central server 22 determines whether the contribution of the first trained model parameter is positive or negative.

[0039] Step S182: If the contribution is a negative value, the central server replaces the federation model with the first contribution model to create the updated federation model.

[0040] As described above, if the contribution calculation module 222 of the central server 22 calculates that the individual or collective contribution of at least one first trained model parameter is negative, it means that adding that at least one first trained model parameter to the federated model will actually decrease the value of the federated model. Therefore, the central server 22 makes the first contribution model obtained by an aggregation algorithm from the remaining trained model parameters after excluding at least one first trained model parameter into a new federated model, thereby enabling the federated model to have higher value and higher accuracy.

[0041] Typically, artificial intelligence models undergo periodic machine learning training on datasets containing both old and new data to ensure continuous updates; this is also true for federative learning. In practice, each AI model training unit can update its trained model by performing machine learning multiple times on its local dataset, and can transmit all the trained model parameters completed in each machine learning run to a central server. The central server then generates the federative model for each machine learning run by running an aggregation algorithm on all the trained model parameters after each training run. For example, each AI model training unit (member of the federative learning) could be configured to update its trained model once every few months. In this case, all AI model training units upload the trained model parameters updated in that run to the central server, and the federative model for that run is generated by running the aggregation algorithm. Therefore, by adjusting the weights of each trained model parameter during the next aggregation algorithm run based on the contribution of each trained model parameter calculated using the above method, a more optimized and accurate federative model can be generated. Furthermore, if a new member joins the federative learning, it can directly join from the next machine learning training and aggregation algorithm stage.

[0042] In practical applications, a federated model is generated by an aggregation algorithm from trained model parameters generated in a single machine learning phase, and then the federated model is redistributed to each federated member (artificial intelligence model training device). Therefore, even if each federated member trains using its own local dataset in the next machine learning phase, updates may be made based on the parameters of the completed federated model. For this reason, when calculating the contribution of the first trained model parameters, each machine learning and aggregation algorithm phase must be taken into consideration. Now, refer to Figure 5. Figure 5 shows a flowchart of the steps for calculating the contribution in federated learning based on another specific embodiment of the present invention. Each step of the method in Figure 5 can be performed by different units of the system in Figure 2. Therefore, below, each step of the method in Figure 5 will be described using the system in Figure 2.

[0043] As shown in Figure 5, the method of this specific embodiment includes the following steps.

[0044] Step S30: The central server 22 accepts multiple trained model parameters. These trained model parameters are generated by performing machine learning multiple times on multiple local datasets (e.g., r times).

[0045] Step S32: The central server 22 generates and stores a federated model by running an aggregation algorithm on all trained model parameters in each machine learning training session.

[0046] Step S340: The central server 22 generates a first contribution model by running an aggregation algorithm on the remaining trained model parameters, excluding the first trained model parameters from each machine learning training session in a simulated manner.

[0047] Step S36: The central server 22 compares the values ​​of the first contribution model and the federated model to generate the contributions of the first trained model parameters.

[0048] In the method of this specific embodiment, when constructing the first contribution model, the federated member that generated the first trained model parameters is excluded from the beginning, and the remaining federated members perform all (r) local and global calculations completely. Therefore, the final first contribution model will be one in which the first trained model parameters do not participate in federated learning at all, and the difference in value between this and the generated federated model will be the contribution of the first trained model parameters. In practice, if, when constructing the first contribution model, both the artificial intelligence model training device 20 and the central server 22 were to revert to their initial state and perform all local and global calculations completely to get as close as possible to the actual contribution, the computational resources and costs would double. Therefore, in this specific embodiment, a method of simulating exclusion (i.e., a method of simulating reverting) is used, but even so, a considerable amount of computational resources and costs may be consumed. This is especially noticeable when calculating the individual contribution for each federated member.

[0049] As mentioned above, if the first trained model parameters are simulated and excluded from the beginning, the resulting contribution will be closest to the actual situation, but it will still consume a significant amount of computational resources and cost. Therefore, in practice, it is possible to determine an appropriate simulation retrospective based on efficiency and computational performance. Now, refer to Figure 6. Figure 6 shows a flowchart of the steps of a method for calculating the contribution in associative learning based on another specific embodiment of the present invention. Each step of the method in Figure 6 can be performed by different units of the system in Figure 2. Therefore, below, each step of the method in Figure 6 will be described using the system in Figure 2. As shown in Figure 6, this specific embodiment differs from the above specific embodiment in the following respects. That is, the method of this specific embodiment further includes the following.

[0050] Step S342: The central server 22 generates a first contribution model by running an aggregation algorithm on the remaining trained model parameters, excluding the first trained model parameters from the last machine learning training.

[0051] In this specific embodiment, the central server 22 can obtain the first contribution model simply by executing an aggregation algorithm after excluding the first trained model parameters from all the last received trained model parameters. In this case, the federation members do not need to simulate backwards or perform any extra machine learning processes. In other words, the method in this specific embodiment requires the central server 22 to execute the aggregation algorithm one more time, resulting in extremely low computational resources and costs. However, since each federation member still bases its calculations on the parameters of the previous federation model in each machine learning stage, the contribution calculated by excluding the first trained model parameters only at the end deviates somewhat from the actual contribution.

[0052] See also Figure 7. Figure 7 shows a flowchart of the steps for calculating contribution in associative learning based on another specific embodiment of the present invention. As shown in Figure 7, this specific embodiment differs from the above specific embodiment in the following respects. That is, the method of this specific embodiment further includes the following.

[0053] Step S344: The central server 22 generates a first contribution model by running an aggregation algorithm on the remaining trained model parameters, excluding the first trained model parameters from the last k-th machine learning training session, counting down from the end. Here, k is an integer between 1 and the number of times the first trained model parameters participated in the aggregation algorithm of the federated model.

[0054] Therefore, the first contribution model can be generated by running an aggregation algorithm on the remaining trained model parameters, excluding the first trained model parameters from the last machine learning stage up to the kth machine learning stage, by counting down from the last machine learning stage. In other words, the process is simulated by counting down from the last machine learning and aggregation algorithm stage up to the kth machine learning and aggregation algorithm stage. Then, the remaining federation members, excluding the federation member that generated the first trained model parameters, and the central server 22 continue to complete the remaining k machine learning and aggregation algorithms, ultimately generating the first contribution model, which excludes the first trained model parameters from the last k machine learning stages. k can be defined as needed. More specifically, the closer the value of k is to the total number of machine learning stages for the first trained model parameters, the closer the calculated contribution will be to the actual contribution, but the more computational resources and costs will be consumed. Conversely, the closer the value of k is to 1, the less computational resources and costs will be consumed, but the more the calculated contribution will deviate from the actual contribution. It should be noted that the other steps in the method of this specific embodiment are almost the same as the corresponding steps in the specific embodiment described above, and therefore will not be described in detail again here.

[0055] In the specific embodiments described above, various methods for simulating the exclusion of the first trained model parameters were presented, and these should be selected according to the actual needs. Simply put, the contribution obtained by excluding the first trained model parameters from the initial machine learning stage approaches the actual contribution, but consumes more computational resources and costs. However, when calculating the contribution, it is also possible to reduce the consumption of computational resources and costs while approaching the actual contribution by using an average or weighted average. Now, refer to Figure 8. Figure 8 shows a flowchart of the steps of a method for calculating the contribution in federative learning based on another specific embodiment of the present invention. Each step of the method in Figure 8 can be performed by different units of the system in Figure 2. Therefore, below, each step of the method in Figure 8 will be described using the system in Figure 2. As shown in Figure 8, the method of this specific embodiment includes the following.

[0056] Step S30: The central server 22 accepts multiple trained model parameters. These trained model parameters are generated by performing machine learning multiple times on multiple local datasets (e.g., r times).

[0057] Step S320: The central server 22 generates and stores multiple federated models by running an aggregation algorithm on all trained model parameters in each machine learning training session.

[0058] Step S346: The central server 22 generates multiple first contribution models by removing the first trained model parameters in each machine learning training session and then running an aggregation algorithm on the remaining trained model parameters.

[0059] Step S360: The central server 22 compares the value of the first contribution model generated after each machine learning aggregation with the value of the federated model generated by each aggregation, and generates contributions for multiple first-trained model parameters.

[0060] Step S362: The average contribution is obtained by averaging or weighting the contributions of all first-trained model parameters, and this is used as the contribution of the first-trained model parameter.

[0061] In this specific embodiment, the central server 22 receives the federated model generated by executing the aggregation algorithm after each machine learning stage, and not only redistributes it to the federated members as the basis for the parameters of the next machine learning stage, but also stores it within the central server 22. Furthermore, after receiving the trained model parameters transmitted from all federated members in each machine learning stage, the central server 22 can generate the first contribution model for that machine learning stage by directly excluding the first trained model parameters for that stage and executing the aggregation algorithm on the remaining trained model parameters. Therefore, by executing the aggregation algorithm twice in each machine learning stage, the contribution of the first trained model parameters for that stage can be calculated, and the average contribution can be obtained by averaging or weighting all the calculated contributions. This average contribution is close to the actual contribution because it takes into account the contribution of the first trained model parameters in each machine learning stage to the federated model. In addition, in this specific embodiment, all federated members do not need to go back or run machine learning again; the central server 22 simply needs to execute the aggregation algorithm one more time in each machine learning stage. Therefore, the extra computing resources and costs consumed are extremely low. Furthermore, there is no need to simulate situations in each machine learning training session where the first trained model parameters are not included, using a retrospective simulation method.

[0062] In practical applications, the step of calculating the contribution of each federated member described above can be achieved by the algorithm in the following specific example. According to the specific example, the method for calculating the contribution in federated learning is to obtain a federated model (global model) by further optimizing the parameters of each trained model (local model) after machine learning using the FedAvg algorithm. In this specific example, the algorithm for calculating the contribution of each federated member has the following steps.

[0063] 1. Define a set N formed by many participants, and obtain the value of the global model of set N through r calculations.

[0064]

number

[0065] It is called that.

[0066] 2. Remove the participants from set M from the participants from set N, and calculate the value of the model obtained from rk calculations.

[0067]

number

[0068] Set it as follows.

[0069] 3. For all i∈N,

[0070]

number

[0071] This calculates the marginal contribution of each participant to the global model (i.e., i represents each participant within N).

[0072]

number

[0073] ) can be calculated.

[0074] It should be noted that while FedAvg was used for calculations in this specific embodiment, in practical applications, variations of FedAvg such as FedOpt, FedProx, and SCAFFOLD may be applied to the present invention.

[0075] In this specific example,

[0076]

number

[0077] This represents the value of the model obtained by calculating it for the set N, and therefore represents the case where all participants participate in associative learning. On the other hand,

[0078]

number

[0079] This represents the case where all participants jointly perform rk iterations of associative learning, then the participants from set M are excluded, and the remaining participants complete the final k iterations of associative learning.

[0080]

number

[0081] This can be defined as the marginal contribution of participant i, as it represents the change in the model's value when participant i is absent. In the above calculation, the case where one participant or a set of participants (|M|) is excluded and participates in k of the r associative learning sessions is used as an estimate of the marginal contribution over the entire process (when the participant does not participate in any of the r associative learning sessions). The value of k can be determined as needed. Generally, the larger the value of k or the closer it is to r, the closer the calculated marginal contribution will be to the result after actual exclusion, but the computational cost will also increase accordingly.

[0082] In this specific example, the marginal contribution of participant i

[0083]

number

[0084] This represents the case where only the last of r associative learning iterations is excluded. That is, it represents the case where k is 1. If computational performance allows, k can be freely adjusted, and may be as large as r. That is, participant i may be excluded from all stages. Therefore, the marginal contribution of participant i calculated by excluding k iterations is:

[0085]

number

[0086] This is the result. Also, participant i is one of the participants in set N. That is, the above marginal contribution algorithm considers the case where only one participant is excluded. If we consider excluding all participants in set M simultaneously, the marginal contribution becomes:

[0087]

number

[0088] This is the result.

[0089]

number

[0090] The calculation uses the rk-th FedAvg calculation as the initial value, and then only NM participants perform the subsequent k FedAvg calculations. Therefore, the larger the value of k, the higher the cost required for the calculation.

[0091] Furthermore, in addition to the method of calculating the final marginal contribution by excluding the k participations described above for one or more participants, the value of the model excluding the participant or the group of participants may be calculated in each associative learning session (total of r sessions), and the values ​​of the models calculated in each associative learning session may be averaged or weighted averaged. The value of the model excluding the participant or the group of participants in each associative learning session is:

[0092]

number

[0093] This is defined as follows. Subsequently, the average value of the model after excluding the participant or the group of participants in r associative learning is:

[0094]

number

[0095] Or, if it is further generalized,

[0096]

number

[0097] This is the result. Note that wj is a weighting coefficient, for example,

[0098]

number

[0099] These include other specific arrays such as increasing, decreasing, or partially setting to 0. Defining them in this way allows for a more comprehensive understanding of each participant's contribution from different angles.

[0100] In summary, the method for calculating contribution in associative learning and the contribution calculation / profit sharing system in associative learning of the present invention can calculate the contribution of each member participating in associative learning to the associative model. More specifically, the contribution model is created by excluding the trained model parameters trained by the member and aggregating the remaining trained model parameters. By comparing the value of the contribution model with that of the associative model, the contribution of the member and their trained model parameters can be estimated. Furthermore, the profit sharing ratio of the member can be determined based on the contribution of the member calculated by the system and method of the present invention. Alternatively, the associative model can be made more accurate by adjusting the weight of the member's trained model parameters in the associative model based on the contribution. Moreover, if the member's trained model parameters have a negative impact on the associative model, it is possible to identify and process this from the contribution calculated by the method and system of the present invention. For example, if the contribution of a member's trained model parameters to the associative model is calculated to be negative, the method and system of the present invention can eliminate the negative impact of that member by excluding the member and their trained model parameters and regenerating the associative model.

[0101] The detailed description of the above preferred specific embodiments is intended to more clearly describe the features and spirit of the invention and does not limit the scope of the invention by the preferred specific embodiments disclosed above. Rather, it is intended that various modifications and equivalent configurations are covered within the claims of the invention. Therefore, the claims of the invention should be interpreted most broadly based on the above description to cover all possible modifications and equivalent configurations. [Explanation of Symbols]

[0102] 2. Contribution Calculation and Profit Sharing System in Associative Learning 20. Artificial Intelligence Model Training Device 22 Central Server 220 Computation Modules 222 Contribution Calculation Module 224 Profit Sharing Module Steps in the process: S10-S16, S17, S180-S182, S30-S36, S340-S346, S360-S362

Claims

1. The central server accepts multiple trained model parameters, and these trained model parameters are generated by machine learning on multiple local datasets, The central server generates a federated model by executing an aggregation algorithm on the trained model parameters and saves the federated model. The central server generates a first contribution model by executing the aggregation algorithm on the trained model parameters, excluding or simulating at least one first trained model parameter from the trained model parameters used to generate the federated model, A method for calculating contributions in federated learning, comprising the steps of: the central server comparing the value of the first contribution model with the value of the federated model to generate contributions for the parameters of the first trained model.

2. The aforementioned trained model parameters are generated by performing multiple machine learning training runs on the local dataset. The central server generates the federated model by executing an aggregation algorithm on the trained model parameters generated in each machine learning training session. The method according to claim 1, wherein the central server generates the first contribution model by running the aggregation algorithm on the trained model parameters, excluding the at least one first trained model parameter from the last machine learning training.

3. The aforementioned trained model parameters are generated by performing multiple machine learning training runs on the local dataset. The central server generates the federated model by executing the aggregation algorithm in each machine learning training session. The method according to claim 1, wherein the central server generates the first contribution model by running the aggregation algorithm on the trained model parameters obtained by simulating the exclusion of at least one first trained model parameter in each machine learning training.

4. The aforementioned trained model parameters are generated by performing multiple machine learning training runs on the local dataset. The central server generates the federated model corresponding to each machine learning session by executing the aggregation algorithm in each machine learning training session. The method according to claim 1, wherein the central server generates a first contribution model corresponding to each machine learning session by executing the aggregation algorithm on the trained model parameters excluding the at least one first trained model parameter in each machine learning session, and the central server generates the average contribution as the contribution of the first trained model parameter by comparing the value of the first contribution model in each machine learning session with the value of the federated model and performing an average or weighted average.

5. The aforementioned trained model parameters are generated by performing multiple machine learning training runs on the local dataset. The central server generates the federated model by executing the aggregation algorithm in each machine learning training session. The method according to claim 1, wherein the central server generates the first contribution model by running the aggregation algorithm on the trained model parameters obtained by simulating the exclusion of at least one first trained model parameter from the machine learning training up to the kth time from the end, where k is an integer between 1 and the number of times the first trained model parameter participated in the aggregation algorithm of the federated model.

6. The central server determines whether the contribution of the first trained model parameter is a positive or negative value. The method according to claim 1, further comprising the step that, if the contribution is a negative value, the central server replaces the federation model with the first contribution model to obtain the updated federation model.

7. Multiple artificial intelligence model training devices that generate multiple trained model parameters by machine learning on multiple local datasets, Includes a central server connected to the artificial intelligence model training device and which accepts the trained model parameters, The aforementioned central server is A computing module that generates a federated model by executing an aggregation algorithm on the aforementioned trained model parameters, A contribution calculation module connected to the calculation module generates a first contribution model by executing the aggregation algorithm on the trained model parameters from which at least one first trained model parameter has been excluded or simulated from the trained model parameters used to generate the federated model, and generates the contribution of the first trained model parameter by comparing the value of the first contribution model with the value of the federated model. A contribution calculation and profit distribution system in federative learning, comprising a profit distribution module connected to the contribution calculation module, which receives the contribution and calculates the profit distribution ratio of the first trained model parameters in the federative model based on the contribution.

8. The system according to claim 7, wherein the artificial intelligence model training device performs multiple machine learning training sessions on the local dataset, the computing module generates the federated model by executing an aggregation algorithm in each machine learning training session, and the contribution calculation module generates the first contribution model by executing the aggregation algorithm on the trained model parameters excluding at least one first trained model parameter from the last machine learning training session.

9. The system according to claim 7, wherein the artificial intelligence model training device performs multiple machine learning training sessions on the local dataset, the computing module generates the federated model by executing an aggregation algorithm in each machine learning training session, and the contribution calculation module generates the first contribution model by executing the aggregation algorithm on the trained model parameters obtained by simulating the exclusion of at least one first trained model parameter in each machine learning training session.

10. The system according to claim 7, wherein the artificial intelligence model training device performs multiple machine learning training sessions on the local dataset, the computing module generates the federated model by executing an aggregation algorithm in each machine learning training session, and the contribution calculation module generates the first contribution model by executing the aggregation algorithm on the trained model parameters obtained by simulating the exclusion of at least one first trained model parameter from the last k machine learning training sessions, counting down from the last session, where k is an integer between 1 and the number of times the first trained model parameter participated in the aggregation algorithm of the federated model.

11. The system according to claim 7, wherein the artificial intelligence model training device performs multiple machine learning training sessions on the local dataset, the computing module generates a federated model corresponding to each machine learning session by executing the aggregation algorithm in each machine learning training session, and the contribution calculation module generates a first contribution model corresponding to each machine learning session by executing the aggregation algorithm on the trained model parameters excluding at least one first trained model parameter in each machine learning training session, and compares the value of the first contribution model in each machine learning session with the value of the federated model, and generates the average contribution as the contribution of the first trained model parameter by averaging or weighted averaging.

Images