A joint model training method, device and system

Abstract

This disclosure relates to the field of machine learning technology, and provides a method, apparatus, and system for joint model training. The method includes: acquiring joint model and joint task bidding information issued by a service platform; training the joint model using pre-set training data, obtaining the output result of each training data point, and calculating the loss value between the output result of each training data point and its label value; selecting target data from the training data based on the loss value, the sampling rate of strongly correlated samples, and a loss threshold; determining the total expected incentive value based on the loss value of the target data, and sending a bidding application to the service platform; upon receiving a notification of winning the bid, training the joint model using the target data, obtaining an iterative joint model, and feeding back the model parameters. This disclosure effectively encourages participants to select datasets related to the joint learning task from their training data for joint training, which not only improves the efficiency of model training but also enhances the performance of the trained model.

Description

Technical Field

[0001] This disclosure relates to the field of machine learning technology, and in particular to a joint model training method, apparatus and system. Background Technology

[0002] In federated learning, the relevance of training data to the federated learning task (related to the performance of the model the participants want to achieve or the business problem to be solved) is a key factor determining the performance of the jointly trained application model. Generally, the higher the relevance of the training data to the federated learning task, the better the performance of the jointly trained application model, and the more closely the application model aligns with the business problem the participants want to solve, i.e., the more closely it meets the participants' needs. Furthermore, using training data with low relevance to the federated learning task or low utility can easily lead to reduced model training efficiency (e.g., slow training convergence, increased training epochs), resulting in higher resource consumption.

[0003] However, most participants typically lack expertise in joint learning, making it difficult for them to identify which training data are more relevant to the joint learning task. Furthermore, due to information asymmetry during model training, it becomes even more challenging to select datasets relevant to the joint learning task. Consequently, it is difficult to improve the efficiency of model training, and the resulting model exhibits poor performance. Summary of the Invention

[0004] In view of this, the present disclosure provides a joint model training method, apparatus and system to solve the problem in the prior art that the participants have difficulty in selecting datasets related to the joint learning task for joint training due to reasons such as lack of professional knowledge or information asymmetry in the model training process, resulting in low model training efficiency and poor performance of the trained model.

[0005] A first aspect of this disclosure provides a joint model training method applied to bidding participants, comprising:

[0006] Obtain the joint model and joint task bidding information issued by the service platform. The joint task bidding information includes the joint task, the sampling rate of strongly correlated samples, the loss threshold, and the number of samples required for each round.

[0007] The joint model is trained using the preset training data to obtain the output of each training data. The training data carries a label value, and the loss value between the output of each training data and its label value is calculated.

[0008] Based on the loss value, the sampling rate of strongly correlated samples, and the loss threshold, target data is selected from the training data. The target data includes strongly correlated data with a high degree of correlation with the joint task and weakly correlated data with a low degree of correlation with the joint task. The number of target data is not less than the number of samples required for each round.

[0009] Determine the total expected incentive value based on the loss value of the target data, and send a bidding application to the service platform. The bidding application includes the expected incentive value.

[0010] When a winning bid notification is received from the service platform, the joint model is trained using the target data to obtain an iterative joint model.

[0011] The iterative model parameters of the iterative joint model are fed back to the service platform to obtain the incentive values.

[0012] A second aspect of this disclosure provides a joint model training method applied to a service platform, comprising:

[0013] Obtain task requirement information, which includes model requirements and business requirements;

[0014] Identify bidding participants that meet business requirements;

[0015] The joint model corresponding to the model requirements and the joint task bidding information are issued to the bidding participants. The joint task bidding information includes the joint task, the sampling rate of strongly correlated samples, the loss threshold and the number of samples required in each round, so that the bidding participants can select target data from their preset training data based on the joint model and the joint task bidding information, and determine the total incentive value.

[0016] When the total expected incentive value received from the bidding participants meets the preset incentive value range, a notification of winning the bid is sent to the bidding participants so that they can use the target data to train the joint model and obtain an iterative joint model.

[0017] Incentive values ​​are distributed to the bidding participants based on the iterative model parameters of the iterative joint model provided by them.

[0018] A third aspect of this disclosure provides a joint model incentive training apparatus, applied to a bidding participant, comprising:

[0019] The information acquisition module is configured to acquire joint model and joint task bidding information issued by the service platform. The joint task bidding information includes joint tasks, sampling rate of strongly correlated samples, loss threshold and number of samples required for each round.

[0020] The computation module is configured to train a joint model using preset training data, obtain the output result of each training data, the training data carries label values, and calculate the loss value between the output result of each training data and its label value.

[0021] The data filtering module is configured to filter target data from the training data based on the loss value, the sampling rate of strongly correlated samples, and the loss threshold. The target data includes strongly correlated data with a high degree of correlation with the joint task and weakly correlated data with a low degree of correlation with the joint task. The number of target data is not less than the number of samples required for each round.

[0022] The bidding module is configured to determine the total expected incentive value based on the loss value of the target data and send a bidding application to the service platform, which includes the expected incentive value.

[0023] The training module is configured to train the joint model using the target data when it receives the winning bid notification information sent by the service platform, so as to obtain an iterative joint model.

[0024] The feedback module is configured to feed back the iterative model parameters of the iterative joint model to the service platform in order to obtain incentive values.

[0025] A fourth aspect of this disclosure provides another joint model training apparatus for use in a service platform, comprising:

[0026] The requirement information acquisition module is configured to acquire task requirement information, which includes model requirements and business requirements.

[0027] The determination module is configured to identify bidding participants that meet business requirements;

[0028] The information delivery module is configured to deliver the joint model corresponding to the model requirements and the joint task bidding information to the bidding participants. The joint task bidding information includes the joint task, the sampling rate of strongly correlated samples, the loss threshold, and the number of samples required in each round, so that the bidding participants can select target data from their preset training data based on the joint model and the joint task bidding information, and determine the total incentive value.

[0029] The bid-winning notification module is configured to send a bid-winning notification to the bid-winning participants when the total expected incentive value received from the bid-winning participants meets the preset incentive value range, so that the bid-winning participants can use the target data to train the joint model and obtain an iterative joint model.

[0030] The compensation distribution module is configured to distribute incentive values ​​to bidding participants based on the iterative model parameters of the iterative joint model provided by the bidding participants.

[0031] A fifth aspect of this disclosure provides a joint model training system, comprising:

[0032] The service platform and the bidding participants who communicate with the service platform;

[0033] The service platform includes a joint model training device as described above;

[0034] Bidding participants include another joint model training device as described above.

[0035] The beneficial effects of this disclosure compared to the prior art include at least the following: obtaining joint model and joint task bidding information issued by the service platform, wherein the joint task bidding information includes the joint task, the sampling rate of strongly correlated samples, the loss threshold, and the number of samples required per round; training the joint model using preset training data, obtaining the output result of each training data, wherein the training data carries a label value, and calculating the loss value between the output result of each training data and its label value; and selecting target data from the training data based on the loss value, the sampling rate of strongly correlated samples, and the loss threshold, wherein the target data includes strongly correlated data with a high degree of correlation with the joint task, and data with a high degree of correlation with the joint task. This method uses weakly correlated data with low relevance to the tasks, ensuring the number of target data is no less than the required sample size for each round. Based on the loss value of the target data, a total expected incentive value is determined, and a bidding application, including the expected incentive value, is sent to the service platform. Upon receiving a notification of successful bidding from the service platform, the joint model is trained using the target data to obtain an iterative joint model. The iterative model parameters of the iterative joint model are fed back to the service platform to obtain an incentive value. This method effectively encourages participants to select datasets relevant to the joint learning task from their training data for joint training, improving both the efficiency and performance of the trained model. Furthermore, this method can, to some extent, increase the incentive for participants to provide more datasets relevant to the joint learning task for joint training, which is beneficial for maintaining long-term sustainable joint learning training. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a schematic diagram of a joint learning architecture according to an embodiment of this disclosure;

[0038] Figure 2 This is a flowchart illustrating a joint model training method provided in an embodiment of this disclosure;

[0039] Figure 3 This is a flowchart illustrating another joint model training method provided in this embodiment of the disclosure;

[0040] Figure 4 This is a schematic diagram of the structure of a joint model training device provided in an embodiment of this disclosure;

[0041] Figure 5 This is a schematic diagram of another joint model training device provided in this embodiment;

[0042] Figure 6 This is a schematic diagram of the structure of a joint model training system provided in an embodiment of this disclosure.

[0043] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0044] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, so as to provide a thorough understanding of the embodiments of this disclosure. However, those skilled in the art will understand that this disclosure may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this disclosure with unnecessary detail.

[0045] Federation learning refers to the comprehensive utilization of multiple AI (Artificial Intelligence) technologies, under the premise of ensuring data security and user privacy, to collaboratively explore the value of data and foster new intelligent business forms and models based on joint modeling. Federation learning has at least the following characteristics:

[0046] (1) Participating nodes control their own data in a weakly centralized joint training mode to ensure data privacy and security in the process of co-creating intelligence.

[0047] (2) In different application scenarios, various model aggregation optimization strategies are established by using screening and / or combination of AI algorithms and privacy-preserving computing to obtain high-level and high-quality models.

[0048] (3) Under the premise of ensuring data security and user privacy, based on multiple model aggregation optimization strategies, obtain methods to improve the performance of the federated learning engine. The performance methods can be improved by solving problems such as parallel computing architecture, information interaction under large-scale cross-domain networks, intelligent perception, and anomaly handling mechanisms.

[0049] (4) Obtain the needs of multiple users in various scenarios, determine the true contribution of each joint participant through a mutual trust mechanism, and allocate incentives accordingly.

[0050] Based on the above approach, an AI technology ecosystem based on collaborative learning can be established, fully leveraging the value of industry data and promoting the implementation of scenarios in vertical fields.

[0051] A joint model training method and apparatus according to an embodiment of the present disclosure will now be described in detail with reference to the accompanying drawings.

[0052] Figure 1 This is a schematic diagram of a joint learning architecture according to an embodiment of this disclosure. Figure 1 As shown, the architecture of joint learning may include a service platform 101 and bidding participants 102, 103 and 104, wherein the service platform 101 includes a server (central node) 1011.

[0053] During the joint learning process, a joint model can be established through server 1011, which then sends the model to bidding participants 102, 103, and 104 with which it has established communication connections. Alternatively, any participant can establish the joint model and upload it to server 1011, which then sends it to other bidding participants with which it has established communication connections. Bidding participants 102, 103, and 104 construct the model based on the downloaded basic structure and model parameters, train the model using local data, obtain updated model parameters, and encrypt and upload the updated model parameters to server 1011. Server 1011 aggregates the model parameters sent by bidding participants 102, 103, and 104 to obtain global model parameters, and then transmits the global model parameters back to bidding participants 102, 103, and 104. Bidding participants 102, 103, and 104 iterate their respective models based on the received global model parameters until the models converge, thus training the models and obtaining the final application model. At this point, service platform 101 can send the application model to the model requester and distribute corresponding incentive values ​​to each bidding participant using the budget provided by the model requester.

[0054] During the joint learning training process, bidding participants 102, 103, and 104 can obtain the joint model and joint task bidding information issued by service platform 101 through downloads or other means. The joint task bidding information includes the joint task, the sampling rate of strongly correlated samples, the loss threshold, and the number of samples required per round. They then train the joint model using pre-set training data, obtaining the output result for each training data point. The training data carries label values, and the loss value between the output result and the label value of each training data point is calculated. Next, based on the loss value, the sampling rate of strongly correlated samples, and the loss threshold, target data is selected from the training data. Afterwards, according to... The loss value of the target data determines the total expected incentive value, and a bidding application, including the total expected incentive value, is sent to the service platform. Then, upon receiving the winning bid notification from the service platform, the joint model is trained using the target data to obtain an iterative joint model. The iterative model parameters of the iterative joint model are fed back to the service platform to obtain the incentive value. This can increase the enthusiasm of bidding participants to select training datasets related to the joint task from their training data for joint training, which is conducive to improving the training efficiency and performance of the model. At the same time, the competitive incentive mechanism among the bidding participants can be used to further reduce the total cost of obtaining the application model through joint training.

[0055] It should be noted that the number of bidding participants is not limited to the three mentioned above, but can be set as needed, and this embodiment does not impose any restrictions on this. The aforementioned model requester can be one of the bidding participants who possess training data related to the joint task, or it can be a simple model requester (who does not possess training data related to the joint task).

[0056] Figure 2 This is a flowchart illustrating a joint model training method provided in an embodiment of this disclosure. Figure 2 The joint model training method can be derived from Figure 1 The bidding participants (such as bidding participant 102) shall execute the order. Figure 2 As shown, the joint model training method includes:

[0057] Step S201: Obtain the joint model and joint task bidding information issued by the service platform. The joint task bidding information includes the joint task, the sampling rate of strongly correlated samples, the loss threshold, and the number of samples required for each round.

[0058] Bidding participants typically refer to those who possess training data related to the joint mission.

[0059] In the joint learning process, it typically starts with a randomly initialized model M1 and iterates through n = 1 to N rounds of joint learning until a preset convergence condition is met (e.g., the model accuracy reaches a preset value, or the number of training rounds reaches a preset number, etc.), to obtain the final application model. For example, in the first round of training, service platform 101 can distribute model M1 to each bidding participant; in the second round of training, service platform 101 distributes model M2 to each bidding participant, and so on. In the Nth round of training, service platform 101 distributes model Mn (joint model) to each bidding participant.

[0060] A joint task typically refers to a task related to the business needs of the model requester. For example, if the business need of the model requester is to improve the accuracy of facial recognition in an attendance system, then the joint task is the facial recognition task.

[0061] Strongly correlated samples refer to samples that are highly relevant to the joint task. For example, in a face recognition task, samples highly correlated with the task are images / videos containing faces. The strongly correlated sample sampling rate refers to the percentage of images / videos containing faces extracted from the training samples. Assuming there are X training samples, extracting 15% of the images / videos containing faces from the training samples yields the strongly correlated sample sampling rate.

[0062] The number of samples required per round specifically refers to the number of training samples that each bidding participant needs to contribute to the joint model for training in each round of training.

[0063] The loss threshold can be the upper / lower limit of any loss function, such as the upper / lower limit of cross-entropy loss, mean squared error loss, etc.

[0064] Step S202: Train the joint model using the preset training data, obtain the output result of each training data, the training data carries a label value, and calculate the loss value between the output result of each training data and its label value.

[0065] As an example, suppose the joint model obtained from the service platform is model M1, a face recognition model. The joint task in the joint task bidding information is a face recognition task, the strong correlation sample sampling rate is 80%, the loss threshold is the cross-entropy loss threshold T, and the number of samples required per round is 100. Then, after obtaining the above information through downloading, bidder A can use preset training data (such as 200 image samples containing or not containing faces, these image samples carrying label values ​​indicating whether the image contains a face, for example, using the number 1 (label value 1) to represent the presence of a face, and the number 0 (label value 0) to represent the absence of a face) to train model M1 and obtain the output value of each training data point. Then, the loss value between the output value of each training data point and its label value is calculated, i.e., 200 loss values ​​K1, K2…K are obtained. 200 Where K1 corresponds to the loss value of the first training data, K2 corresponds to the loss value of the second training data, and so on, K... 200 This corresponds to the loss value of the 200th training data point.

[0066] Step S203: Based on the loss value, the sampling rate of strongly correlated samples, and the loss threshold, target data is selected from the training data. The target data includes strongly correlated data with a high degree of correlation with the joint task and weakly correlated data with a low degree of correlation with the joint task. The number of target data is not less than the number of samples required for each round.

[0067] Based on the above example, and according to the aforementioned loss values ​​K1, K2…K… 200 With a strong correlation sample sampling rate of 80% and a loss threshold T, 100 target data points (including 80 image samples containing faces and 20 image samples without faces) are selected from 200 training data points.

[0068] Step S204: Determine the total expected incentive value based on the loss value of the target data, and send a bidding application to the service platform. The bidding application includes the total expected incentive value.

[0069] Generally, a higher loss value indicates a higher correlation between the sample and the joint task, making accurate inference more difficult during joint training. This means greater selection difficulty and higher screening costs, thus requiring higher incentive values ​​from bidding participants. Based on these rules, a correspondence between loss values ​​and expected incentive values ​​can be pre-established. Then, the expected incentive value for each target data point is determined based on this correspondence. Finally, by summing the expected incentive values ​​of all target data, the total expected incentive value can be determined.

[0070] Based on the example above, assuming that among the 100 target data points selected above, the expected incentive value for strongly correlated data (i.e., image samples containing faces) with a high degree of relevance to the joint task is 10 per data point, and the expected incentive value for weakly correlated data (i.e., image samples not containing faces) with a low degree of relevance to the joint task is 1 per data point, then the total expected incentive value can be calculated using the following formula: Total expected incentive value = 10 * 80 + 1 * 20 = 820.

[0071] As an example, the bidding application sent by the bidding participant to the service platform 101 includes a total incentive value of 820 and the average loss value of 80 strongly correlated data points out of the 100 target data points, for the service platform to refer to and determine whether to select the bidding participant to join the joint training in this round.

[0072] Step S205: When the winning bid notification information is received from the service platform, the joint model is trained using the target data to obtain an iterative joint model.

[0073] The notification of winning the bid can specifically be a text message or voice message informing the bidding participants whether they have been selected to participate in this round of joint training. For example, it could be a text message such as "Congratulations on being selected to participate in the first round of joint training".

[0074] As an example, service platform 101 can send the winning bid notification information to the winning bidder by publishing basic information such as the name of the winning bidder on the platform, or by sending the winning bid notification information via instant messaging methods such as SMS or email, so that the winning bidder can determine whether to use the target data it has selected to train the above joint model and obtain an iterative joint model.

[0075] Step S206: Feed back the iterative model parameters of the iterative joint model to the service platform to obtain the incentive value.

[0076] Referring to the example above, once the winning bidder receives the notification of award from the service platform and determines to use the selected target data to train the joint model, obtaining the iterative joint model, they can return the model parameters (e.g., updated weights and biases) to the service platform to receive an incentive value from the platform. This incentive value can refer to an incentive reward, incentive points, etc., that is equal to or unequal to the bidder's total expected incentive value.

[0077] Typically, when a bidding participant receives a notification of winning the bid from the service platform, it means that the service platform accepts the total expected incentive value previously submitted by the bidding participant, and is essentially "committing" to provide the bidding participant with an incentive value no less than that total expected incentive value.

[0078] In addition, a conversion factor between the total expected incentive value and the incentive value can be preset, for example, 1 expected incentive value = 1 incentive value, or 1 expected incentive value = 1.2 incentive values, etc. This conversion factor can be determined through negotiation between the service platform and the bidding participants; no specific restrictions are imposed here.

[0079] The technical solution provided in this disclosure involves obtaining joint model and joint task bidding information from a service platform. The joint task bidding information includes the joint task, the sampling rate of strongly correlated samples, the loss threshold, and the number of samples required per round. The joint model is trained using preset training data to obtain the output result of each training data point. The training data carries label values, and the loss value between the output result and the label value of each training data point is calculated. Based on the loss value, the sampling rate of strongly correlated samples, and the loss threshold, target data is selected from the training data. Target data includes strongly correlated data with a high degree of correlation to the joint task and weakly correlated data with a low degree of correlation to the joint task. The number of target data points is not less than the number of samples required per round. The total expected incentive value is determined based on the loss value of the target data, and a bidding application is sent to the service platform, including the expected incentive value. When a winning bid notification is received from the service platform, the joint model is trained using the target data to obtain an iterative joint model. The iterative model parameters of the iterative joint model are fed back to the service platform to obtain the incentive value. This effectively helps participants select datasets related to the joint learning task from their training data for joint training, improving both the efficiency of model training and the performance of the obtained model. Furthermore, this method can, to some extent, increase the incentive for participants to provide more datasets related to the joint learning task for joint training, which is conducive to maintaining long-term sustainable joint learning training.

[0080] In some embodiments, step S203 above includes:

[0081] Data in the training data whose loss value is greater than or equal to the loss threshold are identified as strongly correlated data with a high degree of correlation to the joint task.

[0082] Data in the training data with a loss value less than the loss threshold are identified as weakly correlated data with a low degree of correlation to the joint task.

[0083] Based on the sampling rate of strongly correlated samples and the number of samples required per round, the first data volume of strongly correlated data and the second data volume of weakly correlated data are determined respectively. The sum of the first data volume and the second data volume is not less than the number of samples required per round.

[0084] Based on the example of step S203 above, these 200 training data points can be numbered, resulting in training data X1 to X200. Assuming that training data X1 to X105 have a loss value greater than or equal to the cross-entropy loss threshold T, then training data X1 to X105 can be identified as strongly correlated data with a high degree of association with the joint task (face recognition task). The remaining training data X106 to X200 have a loss value less than the cross-entropy loss threshold T, then training data X106 to X200 can be identified as weakly correlated data with a low degree of association with the joint task (face recognition task).

[0085] Next, according to the strong correlation sample sampling rate (i.e., 80%), 100*80%=80 data points (i.e., the first data volume) are randomly selected from the training data X1 to X105 as the strong correlation data, and 100*(1-20%)=20 data points (i.e., the second data volume) are randomly selected from the training data X106 to X200 as the weak correlation data.

[0086] In this embodiment of the disclosure, the sampling rate of strongly correlated samples is a decimal between 0 and 1 (excluding 0 and 1), and strongly correlated data and weakly correlated data are extracted according to the above extraction rules. This makes the extracted target data more consistent with the actual sample distribution rules (i.e., including image samples containing faces and image samples not containing faces), which is beneficial to improving the generalization ability of the model obtained through joint training.

[0087] In some embodiments, the step of determining the desired incentive value based on the loss value of the target data specifically includes:

[0088] Based on the loss value of each target data and the pre-defined correspondence between the loss value gradient and the excitation coefficient, the excitation coefficient of each target data is determined.

[0089] The total expected incentive value is determined based on the incentive coefficient of each target data and the preset incentive base.

[0090] As an example, the correspondence between the preset loss gradient and the activation coefficient is shown in Table 1 below.

[0091] Table 1. Correspondence between loss value gradient and excitation coefficient

[0092] Loss gradient Incentive coefficient ≤T ＜1 ＞T ≥1

[0093] Based on the above example, assuming that the strongly correlated data in the target data are training data X1 to X80, and the weakly correlated data are training data X180 to X200, and the loss values ​​of training data X1 to X80 are all greater than T, then their corresponding first incentive coefficients are all ≥1 (assumed to be 1), and the loss values ​​of training data X180 to X200 are all less than T, then their corresponding second incentive coefficients are all <1 (assumed to be 0.5), and the incentive base is set to 100.

[0094] Next, the total expected incentive value of this batch of target data can be calculated according to the following formula: Total expected incentive value = Expected incentive value of strongly correlated data + Expected incentive value of weakly correlated data = Data volume of strongly correlated data * First incentive coefficient * Incentive base + Data volume of weakly correlated data * Second incentive coefficient * Incentive base = 80 * 1 * 100 + 20 * 0.5 * 100 = 9000.

[0095] Figure 3 This is a flowchart illustrating another joint model training method provided in this embodiment. Figure 3 The joint model training method can be derived from Figure 1 The service platform 101 executes. For example... Figure 3 As shown, the joint model training method includes:

[0096] Step S301: Obtain task requirement information, which includes model requirements and business requirements.

[0097] As an example, model requesters can send task requirements to the service platform through reporting or other means. These requirements typically include model requirements and business requirements. Model requirements specify the desired model functionality, required accuracy, and other concrete specifications. Business requirements are usually related to the business problem the model requester is addressing. For example, they might specify the type and quantity of samples needed based on the specific business problem. For instance, to improve the accuracy of face recognition, the model requester might specify that the required samples are images / videos / pictures containing faces, with a minimum of 2000 samples (assuming one image per sample, then 2000 images are needed).

[0098] Step S302: Identify bidding participants that meet the business requirements.

[0099] As a preferred embodiment, based on the task requirements information provided by the model requester, participants with training data that may be suitable for the business needs of the model requester can be selected from a large number of participants covering various industries. This can help the model requester find participants with suitable and efficient training data to participate in joint training more quickly and efficiently, thereby obtaining a model that fits the model requester's needs and improving the training efficiency and performance of the model. At the same time, it can also save various costs in finding suitable bidding participants.

[0100] For example, suppose the business requirement of the model requester is to improve the accuracy of facial recognition, and the specified sample is an image, preferably an image sample containing a face. The service platform can collect the basic information of each participant in advance (including name, ID, industry, data type of sample, location, etc.) and establish a reference table of participants and their basic information as shown in Table 2 below.

[0101] Table 2. Comparison of Participants and Their Basic Information

[0102]

[0103]

[0104] Based on Table 2 above, the service platform can initially screen out participants (such as participants 1 and 2) who have sample data that match the sample type specified by the model requester according to their business needs, preferably images containing human faces, and identify these participants as bidding participants.

[0105] Step S303: Issue the joint model corresponding to the model requirements and the joint task bidding information to the bidding participants. The joint task bidding information includes the joint task, the sampling rate of strongly correlated samples, the loss threshold, and the number of samples required in each round, so that the bidding participants can select target data from their preset training data based on the joint model and the joint task bidding information, and determine the total expected incentive value.

[0106] As an example, the mapping between model requirements and joint models can be pre-defined. Then, based on the model requirements provided by the requesting party, the corresponding joint model is retrieved and distributed to the bidding participants. Simultaneously, the joint task bidding information is also distributed to each bidding participant. After receiving the joint model and joint task bidding information, each bidding participant can follow the steps described above to select target data from their pre-set training data and determine the total incentive value.

[0107] Step S304: When the total expected incentive value received from the bidding participants meets the preset incentive value range, a winning bid notification is sent to the bidding participants so that they can use the target data to train the joint model and obtain an iterative joint model.

[0108] As an example, after each bidding participant selects target data from its training data and determines the total incentive value, it reports the total expected incentive value to the service platform. Upon receiving the total expected incentive values ​​from each bidding participant, the service platform can compare each total expected incentive value with a preset incentive value range (which can be set according to actual conditions, for example, ≤8000 / 100 data points or ≤8500 / 100 data points, etc., without specific limitations here), and then obtain the comparison result.

[0109] For example, suppose five bidders, A, B, C, D, and E, have submitted their total expected incentive values ​​to the service platform, which are 8200 / 100 data points, 8300 / 100 data points, 7500 / 100 data points, 9500 / 100 data points, and 9100 / 100 data points, respectively. The preset incentive value range is ≤8500 / 100 data points. Bidders A, B, and C fall within this range. In this case, a notification of winning the bid can be sent to bidders A, B, and C, allowing them to use their selected target data to train the joint model, resulting in an iterative joint model.

[0110] In this embodiment of the disclosure, by further determining whether the total expected incentive value fed back by each bidding participant is within the preset incentive value range, it is possible to further screen out the participants who have sample data that meet the business needs of the model demanders, and further control the incentive value within a reasonable range, that is, neither too high nor too low. This will not reduce the training enthusiasm of each bidding participant, but will also help the model demanders save the cost of acquiring the required model. In addition, this method can create a healthy competitive atmosphere, so that joint learning can operate healthily and sustainably in the long term.

[0111] Step S305: Based on the iterative model parameters of the iterative joint model provided by the bidding participants, incentive values ​​are issued to the bidding participants.

[0112] The technical solution provided in this disclosure involves acquiring task requirement information, including model requirements and business requirements; identifying bidding participants that meet the business requirements; distributing a joint model corresponding to the model requirements and joint task bidding information to the bidding participants, including the joint task, strong correlation sample sampling rate, loss threshold, and the number of samples required per round, so that the bidding participants can select target data from their preset training data based on the joint model and joint task bidding information, and determine the total incentive value; when the total expected incentive value received from the bidding participants meets the preset incentive value range, sending a winning bid notification to the bidding participants so that they can use the target data to train the joint model and obtain an iterative joint model; and issuing incentive values ​​to the bidding participants based on the iterative model parameters of the iterative joint model provided by the bidding participants. This method effectively encourages participants to select datasets related to the joint learning task from their training data for joint training, improving not only the efficiency of model training but also the performance of the trained model. Furthermore, this method can, to a certain extent, increase the enthusiasm of participants to provide more datasets related to the joint learning task for joint training, which is conducive to maintaining long-term sustainable joint learning training.

[0113] In some embodiments, the aforementioned business requirements include the types of samples required for joint training. Specifically, bidders meeting the business requirements can be identified by following these steps.

[0114] Determine the types of training data possessed by all participants;

[0115] The sample types and training data types are compared to obtain the comparison results. Based on the comparison results, the bidding participants are determined.

[0116] As an example, information such as the name, ID, and sample data (including sample type and quantity) of all registered users (participants) on the service platform can be collected to create a table showing the correspondence between participants and their basic information, as shown in Table 2 above. Then, upon receiving task requirements from the model requester, the specific sample data requirements in the business requirements of the task requirements can be used to find matching participants based on Table 2 above, and these participants can be identified as bidding participants.

[0117] For example, suppose the business requirement of the model requester is to improve the accuracy of face recognition, and the specified samples are images, preferably images containing faces. If, through the query in Table 2 above, participants 1 and 2 are found to match the training data type specified by the model requester (for example, if the description of the sample data type listed in the table is exactly the same as the sample type description specified by the model requester, or if the main keywords are the same, then they are considered a match), then participants 1 and 2 can be identified as bidding participants.

[0118] In some embodiments, step S304 above includes:

[0119] When the total expected incentive value meets the preset incentive value range, obtain the correlation score between the target data provided by the bidding participants and the joint task;

[0120] Based on the correlation score and the total expected incentive value, target participants are selected from the bidding participants, and the winning bid notification information is sent to the target participants so that they can use the target data to train the joint model and obtain an iterative joint model.

[0121] The correlation score is the average loss value of the strongly correlated data in the target data selected by the bidding participants from their training data.

[0122] As an example, target participants are selected from the bidding participants based on the relevance score and the total expected incentive value. Specifically, this includes: calculating the ratio coefficient between the relevance score of each bidding participant and its total expected incentive value; and selecting at least two target participants from the bidding participants based on the ratio coefficient.

[0123] As an example, assume that the total expected incentive values ​​of bidding participants A, B, and C are W1, W2, and W3, respectively, and that W1, W2, and W3 all fall within a preset incentive value range. Their relevance scores are obtained as T1, T2, and T3, respectively. The ratios of the relevance scores of bidding participants A, B, and C to their total expected incentive values ​​are calculated as W1 / T1, W2 / T2, and W3 / T3, respectively. The order of these three ratios from highest to lowest is: W1 / T1 > W3 / T3 > W2 / T2. Furthermore, at least two participants can be selected as target participants based on this ranking from highest to lowest.

[0124] The technical solution provided in this disclosure performs a first screening of bidding participants based on the total expected incentive value reported by the bidding participants. Bidding participants whose expected incentive values ​​meet the preset incentive value range are selected. Then, the correlation score of each bidding participant selected in the first screening is obtained, and the ratio coefficient of their correlation score to the total expected incentive value is calculated. These ratio coefficients are then sorted. Finally, a second screening is performed on these bidding participants based on the sorting results. Bidding participants with total expected incentive values ​​within a suitable range and high correlation scores are selected as target participants. The winning bid notification information is sent to the target participants so that they can use the target data selected by themselves to train the joint model. This not only accelerates the convergence speed of the model but also helps to improve the performance of the model.

[0125] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0126] The following are embodiments of the apparatus disclosed herein, which can be used to execute embodiments of the method disclosed herein. For details not disclosed in the apparatus embodiments of this disclosure, please refer to the embodiments of the method disclosed herein.

[0127] Figure 4 This is a schematic diagram of a joint model training device provided in an embodiment of this disclosure. Figure 4 As shown, the joint model training device includes:

[0128] The information acquisition module 401 is configured to acquire the joint model and joint task bidding information issued by the service platform. The joint task bidding information includes the joint task, the sampling rate of strongly correlated samples, the loss threshold, and the number of samples required for each round.

[0129] The calculation module 402 is configured to train a joint model using preset training data, obtain the output result of each training data, the training data carrying label values, and calculate the loss value between the output result of each training data and its label value.

[0130] The data filtering module 403 is configured to filter target data from the training data based on the loss value, the sampling rate of strongly correlated samples, and the loss threshold. The target data includes strongly correlated data with a high degree of correlation with the joint task and weakly correlated data with a low degree of correlation with the joint task. The number of target data is not less than the number of samples required for each round.

[0131] The bidding module 404 is configured to determine the total expected incentive value based on the loss value of the target data and send a bidding application to the service platform, the bidding application including the total expected incentive value;

[0132] Training module 405 is configured to train the joint model using the target data when it receives the winning bid notification information sent by the service platform, so as to obtain an iterative joint model.

[0133] Feedback module 406 is configured to feed back the iterative model parameters of the iterative joint model to the service platform in order to obtain incentive values.

[0134] The technical solution provided in this disclosure involves an information acquisition module 401 acquiring joint model and joint task bidding information issued by a service platform. The joint task bidding information includes the joint task, the sampling rate of strongly correlated samples, the loss threshold, and the number of samples required per round. A calculation module 402 trains the joint model using preset training data, obtaining the output result of each training data point. The training data carries label values, and the module calculates the loss value between the output result of each training data point and its label value. A data filtering module 403 filters target data from the training data based on the loss value, the sampling rate of strongly correlated samples, and the loss threshold. The target data includes strongly correlated data with a high degree of correlation to the joint task, as well as data with a low degree of correlation to the joint task. For weakly correlated data with low connectivity, the number of target data points should not be less than the number of samples required for each round. The bidding module 404 determines the total expected incentive value based on the loss value of the target data and sends a bidding application to the service platform, which includes the expected incentive value. The training module 405, upon receiving the winning bid notification from the service platform, trains the joint model using the target data to obtain an iterative joint model. The feedback module 406 feeds back the iterative model parameters of the iterative joint model to the service platform to obtain an incentive value. This method effectively encourages participants to select datasets related to the joint learning task from their training data for joint training, improving both the efficiency and performance of the trained model. Furthermore, this method can, to some extent, increase the enthusiasm of participants to provide more datasets related to the joint learning task for joint training, which is beneficial for maintaining long-term sustainable joint learning training.

[0135] In some embodiments, the data filtering module 403 includes:

[0136] The first determining unit is configured to determine data in the training data whose loss value is greater than or equal to the loss threshold as strongly correlated data with a high degree of correlation with the joint task.

[0137] The second determining unit is configured to determine data in the training data whose loss value is less than the loss threshold as weakly correlated data with low correlation to the joint task.

[0138] The data volume determination unit is configured to determine the first data volume of strongly correlated data and the second data volume of weakly correlated data based on the sampling rate of strongly correlated samples and the number of samples required in each round, wherein the sum of the first data volume and the second data volume is not less than the number of samples required in each round.

[0139] In some embodiments, determining the desired incentive value based on the loss value of the target data includes:

[0140] Based on the loss value of each target data and the pre-defined correspondence between the loss value gradient and the excitation coefficient, the excitation coefficient of each target data is determined.

[0141] The total expected incentive value is determined based on the incentive coefficient of each target data and the preset incentive base.

[0142] Figure 5 This is a schematic diagram of a joint model training device provided in an embodiment of this disclosure. Figure 5 As shown, the joint model training device includes:

[0143] The requirement information acquisition module 501 is configured to acquire task requirement information, which includes model requirements and business requirements.

[0144] Module 502 is configured to identify bidding participants that meet business requirements;

[0145] The information delivery module 503 is configured to deliver the joint model corresponding to the model requirements and the joint task bidding information to the bidding participants. The joint task bidding information includes the joint task, the sampling rate of strongly correlated samples, the loss threshold and the number of samples required in each round, so that the bidding participants can select target data from their preset training data based on the joint model and the joint task bidding information, and determine the total incentive value.

[0146] The bid-winning notification module 504 is configured to send a bid-winning notification to the bid-winning participants when the total expected incentive value received from the bid-winning participants meets the preset incentive value range, so that the bid-winning participants can use the target data to train the joint model and obtain an iterative joint model.

[0147] The distribution module 505 is configured to distribute incentive values ​​to the bidding participants based on the iterative model parameters of the iterative joint model provided by the bidding participants.

[0148] The technical solution provided in this disclosure involves: a requirement information acquisition module 501 acquiring task requirement information, including model requirements and business requirements; a determination module 502 identifying bidding participants that meet the business requirements; an information distribution module 503 distributing a joint model corresponding to the model requirements and joint task bidding information to the bidding participants, including the joint task, strong correlation sample sampling rate, loss threshold, and the number of samples required per round, so that the bidding participants can select target data from their preset training data based on the joint model and joint task bidding information, and determine the total incentive value; a winning bid notification module 504 sending a winning bid notification to the bidding participants when the total expected incentive value received from the bidding participants meets the preset incentive value range, so that the bidding participants can use the target data to train the joint model and obtain an iterative joint model; and an incentive distribution module 505 distributing incentive values ​​to the bidding participants based on the iterative model parameters of the iterative joint model provided by the bidding participants. This effectively encourages participants to select datasets related to the joint learning task from their training data for joint training, which not only improves the efficiency of model training but also enhances the performance of the trained model. Furthermore, this method can, to some extent, increase the incentive for participants to provide more datasets related to the joint learning task for joint training, which is conducive to maintaining long-term sustainable joint learning training.

[0149] In some embodiments, the aforementioned business requirements include the sample types required for joint training. The aforementioned determining module 502 includes:

[0150] The type determination unit is configured to determine the type of training data possessed by all participants;

[0151] The bidding unit is configured to compare sample types and training data types, obtain comparison results, and determine bidding participants based on the comparison results.

[0152] In some embodiments, the above-mentioned bid-winning notification module 504 includes:

[0153] The scoring acquisition unit is configured to acquire a score on the correlation between the target data provided by the bidding participants and the joint task when the total expected incentive value meets the preset incentive value range.

[0154] The bid-winning notification unit is configured to select target participants from the bidding participants based on the correlation score and the total expected incentive value, and send bid-winning notification information to the target participants so that they can use the target data to train the joint model and obtain an iterative joint model.

[0155] In some embodiments, selecting target participants from bidding participants based on the total expected incentive value of the relevance score includes:

[0156] Calculate the ratio of the relevance score of each bidding participant to its total expected incentive value;

[0157] Based on the ratio coefficient, at least two target participants are selected from the bidding participants.

[0158] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this disclosure.

[0159] Figure 6 This is a schematic diagram of the structure of a joint model training system provided in an embodiment of this disclosure. Figure 6 As shown, the joint model training system includes a service platform 101 and bidding participants 102 and model requesters 601, which are communicatively connected to the service platform. The service platform 101 includes a joint model training device as shown in Figure 5, and the bidding participants 102 include... Figure 4 The joint model training device shown.

[0160] Specifically, when model requester 601 needs to use an application model to solve a certain business problem, it can send task request information to service platform 101 via network, Bluetooth, or other means. After receiving the task request information, service platform 101 can determine the bidding participant 102 that meets the business requirements based on the task request information, and then issue the joint model corresponding to the model requirement, as well as the joint task bidding information, to the bidding participant 102. The joint task bidding information includes the joint task, the sampling rate of strongly correlated samples, the loss threshold, and the number of samples required per round. At this time, after receiving the above-mentioned joint model and joint task bidding information, bidding participant 102 selects target data from its preset training data according to the joint model and joint task bidding information, determines the total expected incentive value, and applies the total expected incentive value. The incentive value is fed back to the service platform 101. When the service platform 101 receives the total expected incentive value from the bidding participant 102 and it meets the preset incentive value range, it sends a winning bid notification to the bidding participant 102. After receiving the winning bid notification from the service platform 101, the bidding participant 102 uses the target data it has selected to train the joint model Mn, obtains an iterative joint model, and feeds back the model parameters of the iterative joint model to the service platform 101. After receiving the model parameters of the iterative joint model from each bidding participant 102, the service platform 101 aggregates the model parameters from each bidding participant and updates the joint model using the aggregated parameters to obtain the joint model M(n+1). The service platform 101 then issues the corresponding incentive value to each bidding participant, completing one round of update and iterative training. The above steps can then be repeated until the joint model reaches the preset convergence condition, obtaining the final application model. The application model is then returned to the model requester, and the incentive values ​​paid to each bidding participant, as well as other miscellaneous expenses, are deducted from the budget provided by the model requester 601.

[0161] The technical solution provided in this disclosure effectively encourages participants to select datasets related to the joint learning task from their training data for joint training. This not only improves the efficiency of model training but also enhances the performance of the trained model. Furthermore, this method can, to some extent, increase the incentive for participants to provide more datasets related to the joint learning task for joint training, which is beneficial for maintaining long-term sustainable joint learning training.

[0162] Figure 7 This is a schematic diagram of the electronic device 700 provided in an embodiment of this disclosure. Figure 7As shown, the electronic device 700 of this embodiment includes a processor 701, a memory 702, and a computer program 703 stored in the memory 702 and executable on the processor 701. When the processor 701 executes the computer program 703, it implements the steps in the various method embodiments described above. Alternatively, when the processor 701 executes the computer program 703, it implements the functions of each module / unit in the various device embodiments described above.

[0163] For example, computer program 703 may be divided into one or more modules / units, which are stored in memory 702 and executed by processor 701 to perform the present disclosure. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 703 in electronic device 700.

[0164] Electronic device 700 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 700 may include, but is not limited to, a processor 701 and a memory 702. Those skilled in the art will understand that... Figure 7 This is merely an example of electronic device 700 and does not constitute a limitation on electronic device 700. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device may also include input / output devices, network access devices, buses, etc.

[0165] The processor 701 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0166] The memory 702 can be an internal storage unit of the electronic device 700, such as a hard disk or RAM of the electronic device 700. The memory 702 can also be an external storage device of the electronic device 700, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the electronic device 700. Furthermore, the memory 702 can include both internal and external storage units of the electronic device 700. The memory 702 is used to store computer programs and other programs and data required by the electronic device. The memory 702 can also be used to temporarily store data that has been output or will be output.

[0167] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0168] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0169] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0170] In the embodiments provided in this disclosure, it should be understood that the disclosed devices / electronic devices and methods can be implemented in other ways. For example, the device / electronic device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. Multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0171] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0172] Furthermore, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0173] If an integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in a computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0174] The above embodiments are only used to illustrate the technical solutions of this disclosure, and are not intended to limit it. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be included within the protection scope of this disclosure.

Claims

1. A joint model training method, characterized in that, include: Obtain the joint model and joint task bidding information issued by the service platform. The joint task bidding information includes the joint task, the sampling rate of strongly correlated samples, the loss threshold, and the number of samples required for each round. Among them, the joint task includes the target object recognition task; strongly correlated samples refer to samples with high correlation to the joint task; the sampling rate of strongly correlated samples refers to the percentage of images or videos containing the target object extracted from the training samples. The joint model is trained using preset training data to obtain the output result of each training data, wherein the training data carries a label value, and the loss value between the output result of each training data and its label value is calculated. Based on the loss value, the sampling rate of strongly correlated samples, and the loss threshold, target data is selected from the training data. The target data includes strongly correlated data with a high degree of correlation with the joint task and weakly correlated data with a low degree of correlation with the joint task. The number of target data is not less than the number of samples required for each round. Based on the loss value of the target data, a total expected incentive value is determined, and a bidding application is sent to the service platform, the bidding application including the total expected incentive value; When the winning bid notification information is received from the service platform, the target data is used to train the joint model to obtain an iterative joint model; The iterative model parameters of the iterative joint model are fed back to the service platform to obtain the incentive value.

2. The method according to claim 1, characterized in that, The step involves selecting target data from the training data based on the loss value, the sampling rate of strongly correlated samples, and the loss threshold. The target data includes strongly correlated data with a high degree of correlation to the joint task, and weakly correlated data with a low degree of correlation to the joint task. The number of target data is not less than the number of samples required for each round, including: Data in the training data whose loss value is greater than or equal to the loss threshold are identified as strongly correlated data with a high degree of correlation to the joint task. Data in the training data whose loss value is less than the loss threshold are identified as weakly correlated data with low correlation to the joint task. Based on the sampling rate of the strongly correlated samples and the number of samples required for each round, the first data volume of the strongly correlated data and the second data volume of the weakly correlated data are determined respectively, and the sum of the first data volume and the second data volume is not less than the number of samples required for each round.

3. The method according to claim 1, characterized in that, The step of determining the total expected incentive value based on the loss value of the target data includes: Based on the loss value of each target data and the pre-defined correspondence between the loss value gradient and the excitation coefficient, the excitation coefficient of each target data is determined. The total expected incentive value is determined based on the incentive coefficient of each target data and the preset incentive base.

4. A joint model training method, characterized in that, include: Obtain task requirement information, which includes model requirements and business requirements; Identify bidding participants who meet the aforementioned business requirements; The bidding participants are issued a joint model corresponding to the model requirements, as well as joint task bidding information. The joint task bidding information includes the joint task, the sampling rate of strongly correlated samples, the loss threshold, and the number of samples required per round. This allows the bidding participants to select target data from their preset training data based on the joint model and the joint task bidding information, and determine the total expected incentive value. The joint task includes a target object recognition task. Strongly correlated samples refer to samples with a high correlation to the joint task. The sampling rate of strongly correlated samples refers to the percentage of images or videos containing the target object extracted from the training samples. When the total expected incentive value received from the bidding participant meets the preset incentive value range, a winning bid notification is sent to the bidding participant so that the bidding participant can use the target data to train the joint model and obtain an iterative joint model. Incentive values ​​are issued to the bidding participants based on the iterative model parameters of the iterative joint model provided by the bidding participants.

5. The method according to claim 4, characterized in that, The business requirements include the types of samples needed for joint training; The process of identifying bidding participants that meet the business requirements includes: Determine the types of training data possessed by all participants; The sample type and the training data type are compared to obtain a comparison result, and the bidding participants are determined based on the comparison result.

6. The method according to claim 4, characterized in that, When the total expected incentive value received from the bidding participant meets the preset incentive value range, a winning bid notification is sent to the bidding participant so that the bidding participant can use the target data to train the joint model and obtain an iterative joint model, including: When the total expected incentive value meets the preset incentive value range, obtain the correlation score between the target data provided by the bidding participants and the joint task; Based on the correlation score and the total expected incentive value, target participants are selected from the bidding participants, and a winning bid notification is sent to the target participants so that they can use the target data to train the joint model and obtain an iterative joint model.

7. The method according to claim 6, characterized in that, The step of selecting target participants from the bidding participants based on the relevance score and the total expected incentive value includes: Calculate the ratio of the relevance score of each bidding participant to its total expected incentive value; Based on the aforementioned ratio coefficient, at least two target participants are selected from the bidding participants.

8. A joint model training device, characterized in that, include: The information acquisition module is configured to acquire joint model and joint task bidding information issued by the service platform. The joint task bidding information includes joint tasks, strong correlation sample sampling rate, loss threshold, and the number of samples required per round. Among them, the joint task includes target object recognition task; strong correlation sample refers to sample with high correlation to the joint task; strong correlation sample sampling rate refers to the percentage of images or videos containing the target object extracted from the training samples. The calculation module is configured to train the joint model using preset training data, obtain the output result of each training data, wherein the training data carries a label value, and calculate the loss value between the output result of each training data and its label value. The data filtering module is configured to filter target data from the training data based on the loss value, the sampling rate of strongly correlated samples, and the loss threshold. The target data includes strongly correlated data with a high degree of correlation with the joint task and weakly correlated data with a low degree of correlation with the joint task. The number of target data is not less than the number of samples required for each round. The bidding module is configured to determine the total expected incentive value based on the loss value of the target data and send a bidding application to the service platform, the bidding application including the total expected incentive value; The training module is configured to train the joint model using the target data when it receives the winning bid notification information sent by the service platform, so as to obtain an iterative joint model. The feedback module is configured to feed back the iterative model parameters of the iterative joint model to the service platform in order to obtain the incentive value.

9. A joint model training device, characterized in that, include: The requirement information acquisition module is configured to acquire task requirement information, which includes model requirements and business requirements. The determination module is configured to determine the bidding participants that meet the business requirements; The information delivery module is configured to deliver a joint model corresponding to the model requirements and joint task bidding information to the bidding participants. The joint task bidding information includes the joint task, the strong correlation sample sampling rate, the loss threshold, and the number of samples required per round, so that the bidding participants can select target data from their preset training data based on the joint model and the joint task bidding information, and determine the total expected incentive value. The joint task includes a target object recognition task; strong correlation samples refer to samples with high correlation to the joint task; and the strong correlation sample sampling rate refers to the percentage of images or videos containing the target object extracted from the training samples. The bid-winning notification module is configured to send a bid-winning notification to the bid-winning participant when the total expected incentive value received from the bid-winning participant meets the preset incentive value range, so that the bid-winning participant can use the target data to train the joint model and obtain an iterative joint model. The reward distribution module is configured to distribute incentive values ​​to the bidding participants based on the iterative model parameters of the iterative joint model provided by the bidding participants.

10. A joint model training system, characterized in that, include: The service platform and the bidding participants and model requesters who communicate with the service platform; The service platform includes the joint model training device as described in claim 9; The bidding participants include the joint model training device as described in claim 8.

Images