Method, device, electronic device and storage medium for improving data processing accuracy

Abstract

The present disclosure relates to a method, device, electronic device and storage medium for improving data processing accuracy to solve the problem of low accuracy of business classification, including: determining target data information of a target user; obtaining target features of the target user according to the target data information of the target user based on a called cloud neural network model, wherein the cloud neural network model is trained in the following manner: obtaining multiple sub-models, each of which is trained based on the original data information of a seed user as a training sample, using the multiple sub-models as target training samples of the cloud neural network model, and training the cloud neural network model; the cloud neural network model is used to determine the target features of the target user using the original data information as an input, and output the target features as an output. The data processing accuracy and precision can be improved.

Description

Technical Field

[0001] The present disclosure relates to the field of medical technology, and in particular to a method, device, electronic device and storage medium for improving data processing accuracy. Background Art

[0002] In the era of AI (Artificial Intelligence), big data, as the basis of AI model training, is an indispensable part of the AI ​​model training process. The larger the amount of data, the higher the accuracy of the trained model. Therefore, in the process of model training, a large amount of data often needs to be collected. However, based on the protection of user personal information, the amount of data collected is small, and the user's personal information cannot be unified together for training, so the accuracy of the trained model is low. For example, in the medical industry, relevant original data such as the patient's personal medical records and clinical data cannot leave the hospital or medical data center. In this way, there are fewer medical records and clinical data, and the personal medical records and clinical data between hospitals cannot be unified together. The accuracy of the trained AI model is low, and it is difficult to accurately determine the vertical group of users, such as the patient group, when applying it.

[0003] At present, under the premise of ensuring data security, users' AI sub-models can be uploaded to the cloud and synthesized into larger AI models. For example, a larger AI model can be synthesized based on a decision tree. However, simply synthesizing a larger AI model will result in a complex neural network structure for the new AI model, requiring large computing resources, and deploying it to application scenarios. The conditions required are also more complex, requiring a large investment in computing resources, and the model operation time is also longer.

[0004] Therefore, how to reasonably train AI models to improve data processing accuracy while ensuring the security of relevant original data such as user personal information has become an urgent problem to be solved. Summary of the invention

[0005] In order to overcome the problems existing in the related art, the present disclosure provides a method, device, electronic device and storage medium for improving data processing accuracy.

[0006] According to a first aspect of an embodiment of the present disclosure, a method for improving data processing accuracy is provided, the method comprising:

[0007] Determine target data information of target users;

[0008] Based on the called cloud neural network model, the data features of the target user are obtained according to the target data information of the target user, wherein the cloud neural network model is trained in the following manner: a plurality of sub-models are obtained, each of which is obtained by training based on the original data information of the seed user as a training sample, and the plurality of sub-models are used as target training samples of the cloud neural network model to obtain the cloud neural network model through training;

[0009] The cloud-based neural network model is used to use the target data information as input to determine the target features of the target user, and output the target features as output.

[0010] Preferably, the step of using the multiple sub-models as target training samples of the cloud-based neural network model to train the cloud-based neural network model includes:

[0011] Calculate the complexity of each sub-model with respect to the corresponding training sample to obtain the generalization error corresponding to the sub-model;

[0012] Using the sub-model with a generalization error less than or equal to a preset threshold as the target training sample for training the cloud neural network model;

[0013] The cloud-based neural network model is trained based on the target training samples.

[0014] Preferably, the calculation of the complexity of each sub-model with respect to the corresponding training sample to obtain the generalization error corresponding to the sub-model is specifically as follows:

[0015]

[0016] Among them, L D (h) is the generalization error, L S (h) is the training error, is the Rademacher complexity, is the loss function, is the hypothesis class, S is the number of samples in the training set, and L is the loss function The constant for Lipschitz continuity of the prediction term, r>0 is the radius of the local Rademacher complexity, which is used to constrain the hypothesis class To obtain a tighter upper bound, λ j , j = 1...m is a symmetric square matrix Eigenvalues ​​in non-ascending order, It represents the value of the i-th sample propagated to the layer immediately adjacent to the output layer, σ is the confidence level, and 0<σ<1. Preferably, the confidence level σ is 0.95.

[0017] Preferably, the cloud-based neural network model training further includes:

[0018] Requesting the sub-model for the amount of data of the training samples of the sub-model and the gradient of the training error in the model parameters;

[0019] Calculating the gradient direction of the cloud neural network model according to the data volume of the training samples and the gradient of the training error in the model parameters;

[0020] Based on the gradient descent method, the cloud neural network model is updated according to the gradient direction;

[0021] Determine whether the updated cloud-based neural network model converges, and if the cloud-based neural network model does not converge, recalculate the gradient direction of the cloud-based neural network model.

[0022] Preferably, the gradient direction of the cloud neural network model is calculated as follows:

[0023]

[0024] in, is the gradient direction, is the sample size of the training set.

[0025] According to a second aspect of the embodiments of the present disclosure, there is provided a device for improving data processing accuracy, the device comprising:

[0026] An acquisition module is configured to determine target data information of a target user;

[0027] The execution module is configured to obtain the data features of the target user according to the target data information of the target user based on the called cloud neural network model, wherein the cloud neural network model is trained in the following manner: obtaining a plurality of sub-models, each of which is trained based on the original data information of the seed user as a training sample, using the plurality of sub-models as target training samples of the cloud neural network model, and training to obtain the cloud neural network model;

[0028] The cloud-based neural network model is used to use the target data information as input to determine the target features of the target user, and output the target features as output.

[0029] Preferably, the multiple sub-models are used as target training samples of the cloud neural network model to train the cloud neural network model, specifically:

[0030] Calculate the complexity of each sub-model with respect to the corresponding training sample to obtain the generalization error corresponding to the sub-model;

[0031] Using the sub-model with a generalization error less than or equal to a preset threshold as the target training sample for training the cloud neural network model;

[0032] The cloud-based neural network model is trained based on the target training samples.

[0033] Preferably, the calculation of the complexity of each sub-model with respect to the corresponding training sample to obtain the generalization error corresponding to the sub-model is specifically as follows:

[0034]

[0035] Among them, L D (h) is the generalization error, L S (h) is the training error, is the Rademacher complexity, is the loss function, is the hypothesis class, S is the number of samples in the training set, and L is the loss function The constant for Lipschitz continuity of the prediction term, r>0 is the radius of the local Rademacher complexity, which is used to constrain the hypothesis class To obtain a tighter upper bound, λ j , j = 1...m is a symmetric square matrix Eigenvalues ​​in non-ascending order, It represents the value of the i-th sample propagated to the layer immediately adjacent to the output layer, σ is the confidence level, and 0<σ<1. Preferably, the confidence level σ is 0.95.

[0036] Preferably, the cloud-based neural network model training further includes:

[0037] Requesting the sub-model for the amount of data of the training samples of the sub-model and the gradient of the training error in the model parameters;

[0038] Calculating the gradient direction of the cloud neural network model according to the data volume of the training samples and the gradient of the training error in the model parameters;

[0039] Based on the gradient descent method, the cloud neural network model is updated according to the gradient direction;

[0040] Determine whether the updated cloud-based neural network model converges, and if the cloud-based neural network model does not converge, recalculate the gradient direction of the cloud-based neural network model.

[0041] Preferably, the gradient direction of the cloud neural network model is calculated as follows:

[0042]

[0043] in, is the gradient direction, is the sample size of the training set.

[0044] According to a third aspect of the present disclosure, an electronic device is provided, including:

[0045] A memory storing programmable control instructions;

[0046] A processor is used to execute the programmable control instructions in the memory to implement the steps of any method described in the first aspect.

[0047] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, on which computer program instructions are stored. When the program instructions are executed by a processor, the steps of any one of the methods described in the first aspect are implemented.

[0048] The technical solution provided by the embodiment of the present disclosure may include the following beneficial effects: determining the target data information of the target user; obtaining the data features of the target user according to the target data information of the target user based on the called cloud neural network model, and the neural network model is trained in the following manner: obtaining multiple sub-models, each of which is trained based on the original data information of the seed user as a training sample, and using the multiple sub-models as the target training samples of the cloud neural network model to train the cloud neural network model; the cloud neural network model is used to use the original data information as an input to determine the target features of the target user, and output the target features as an output. The data processing accuracy can be improved, and the accuracy of user classification can be improved. Moreover, the original data information of the seed users of each sub-model can be trained to obtain an accurate cloud neural network model without leaving the data source, which can improve the accuracy of user classification. Moreover, multiple sub-models can be fused into a new model in the cloud, and its structure is consistent with the model structure of the sub-model, without changing the model architecture corresponding to each sub-model, reducing the amount of calculation and saving costs.

[0049] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0051] Figure 1 The present invention is a flowchart of a method for improving data processing accuracy according to an exemplary embodiment.

[0052] Figure 2 The present invention is a flowchart of a neural network model training according to an exemplary embodiment.

[0053] Figure 3 FIG. 4 is a flowchart showing a method for implementing step S122 according to an exemplary embodiment.

[0054] Figure 4 The figure is a flowchart of another neural network model training according to an exemplary embodiment.

[0055] Figure 5 The figure is a flowchart of another neural network model training according to an exemplary embodiment.

[0056] Figure 6 It is a schematic diagram showing a comparison between a training error and a generalization error according to an exemplary embodiment.

[0057] Figure 7 The invention is a block diagram of a device for improving data processing accuracy according to an exemplary embodiment. DETAILED DESCRIPTION

[0058] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0059] Figure 1 is a flow chart of a method for improving data processing accuracy according to an exemplary embodiment. Figure 1 As shown, the method can be applied to medical devices, such as blood pressure monitors, blood glucose monitors, and electronic devices for image processing, etc. The method includes the following steps.

[0060] In step S11, target data information of a target user is determined.

[0061] In step S12, based on the called cloud-based neural network model, the data features of the target user are obtained according to the target data information of the target user.

[0062] The cloud-based neural network model is used to use the target data information as input to determine the target features of the target user, and output the target features as output.

[0063] in, Figure 2 FIG. 1 is a flowchart of a neural network model training according to an exemplary embodiment. Figure 2 As shown, the neural network model is trained in the following manner including the following steps:

[0064] In step S121, a plurality of sub-models are obtained, wherein each of the sub-models is trained based on the original data information of the seed user as a training sample.

[0065] In step S122, the multiple sub-models are used as target training samples of the cloud-based neural network model to train and obtain the cloud-based neural network model.

[0066] The above technical solution determines the target data information of the target user; based on the called cloud neural network model, the data feature value of the target user is obtained according to the target data information of the target user. The neural network model is trained in the following way: multiple sub-models are obtained, each sub-model is trained based on the original data information of the seed user as a training sample, and the multiple sub-models are used as the target training samples of the cloud neural network model to train the cloud neural network model; according to the data feature value of the target user, the target classification of the target user is determined. It can improve the data processing accuracy and the accuracy of user classification. Moreover, the original data information of the seed users of each sub-model can be trained to obtain an accurate cloud neural network model without leaving the data source, which can improve the accuracy of user classification. Moreover, multiple sub-models can be fused into a new model in the cloud, and its structure is consistent with the model structure of the sub-model, without changing the model architecture corresponding to each sub-model, reducing the amount of calculation and saving costs.

[0067] Preferably, Figure 3 The flowchart shown is an exemplary implementation of step S122. Figure 3 As shown, in step S122, the multiple sub-models are used as target training samples of the cloud neural network model to train the cloud neural network model, including:

[0068] In step S1221, the complexity of each sub-model with respect to the corresponding training sample is calculated to obtain the generalization error corresponding to the sub-model;

[0069] In step S1222, the sub-model whose generalization error is less than or equal to a preset threshold is used as the target training sample for training the cloud neural network model;

[0070] In step S1223, the cloud-based neural network model is trained based on the target training sample.

[0071] Optionally, the calculation of the complexity of each sub-model with respect to the corresponding training sample to obtain the generalization error corresponding to the sub-model is specifically:

[0072]

[0073] Among them, L D (h) is the generalization error, L S (h) is the training error, is the Rademacher complexity, is the loss function, is the hypothesis class, S is the number of samples in the training set, and L is the loss function The constant for Lipschitz continuity of the prediction term, r>0 is the radius of the local Rademacher complexity, which is used to constrain the hypothesis class To obtain a tighter upper bound, λ j , j = 1...m is a symmetric square matrix Eigenvalues ​​in non-ascending order, It represents the value of the i-th sample propagated to the layer immediately adjacent to the output layer, σ is the confidence level, and 0<σ<1. Preferably, the confidence level σ is 0.95.

[0074] Preferably, Figure 4 FIG. 1 is a flowchart of a neural network model training according to an exemplary embodiment. Figure 4 As shown, the neural network model training also includes:

[0075] In step S123, the sub-model is requested to obtain the amount of data of the training samples of the sub-model and the gradient of the training error in the model parameters;

[0076] In step S124, the gradient direction of the cloud neural network model is calculated according to the data volume of the training sample and the gradient of the training error in the model parameters;

[0077] In step S125, based on the gradient descent method, the cloud neural network model is updated according to the gradient direction;

[0078] In step S126, it is determined whether the updated cloud-based neural network model has converged, and if the cloud-based neural network model has not converged, the gradient direction of the cloud-based neural network model is calculated again.

[0079] The above technical solution allows the sub-model to communicate with the cloud model to update the cloud model. Local users can quickly switch between the local model and the fused model, and the calculation time will not be extended, reducing the time cost. In addition, the sub-model does not need to upload the original data in training. All the information uploaded by each sub-model during communication cannot be restored to the training data, ensuring the security of hospital patient data.

[0080] Figure 5 is a flowchart of a neural network model training according to an exemplary embodiment. Figure 5 As shown, the neural network model is taken as a medical model as an example. For example, first, m sub-models are trained locally based on the original medical records and original clinical data of the patient to obtain multiple sub-models. It can be explained that the sub-models can be used to determine the probability of the patient's illness.

[0081] Furthermore, the m sub-models obtained through training are obtained, and the complexity of each sub-model with respect to the training data is calculated to obtain the generalization error corresponding to the sub-model, so as to evaluate the sub-model and further determine the target sub-model as the initial value.

[0082] Furthermore, each evaluation result is uploaded to the cloud so that the cloud can select the target sub-model as the initial value.

[0083] Furthermore, a sub-model is selected as the initial value based on the evaluation results, and the medical model is trained.

[0084] Furthermore, it is determined whether the obtained medical model needs to be updated, and if it is determined that the medical model does not need to be updated, a fused model is obtained.

[0085] When it is determined that the medical model needs to be updated, the local algorithm is requested to obtain the data required by the algorithm, which is non-training data. And k feedbacks are received to update the medical model, where 0 <k<m+1。

[0086] Further, it is determined whether the cloud model, namely the medical treatment, has converged. If it is determined that the cloud model has converged, a fused model is obtained. If it is determined that the cloud model has converged, the step of requesting data required by the algorithm from the local is continued.

[0087] In this way, the accuracy of determining the vertical group of patients can be improved, and the accuracy of model training can also be improved.

[0088] Preferably, the calculating of the gradient direction of the neural network model is specifically to obtain the gradient direction by the following analytical formula:

[0089]

[0090] in, is the gradient direction, is the sample size of the training set.

[0091] Figure 6 It is a schematic diagram showing a comparison between a training error and a generalization error according to an exemplary embodiment.

[0092] Figure 6 There are 6 subgraphs in the figure, each of which represents a submodel. The curve where the triangle is located represents the training data set of the submodel, the curve where the triangle is located represents the actual data distribution (the distribution is unknown in reality), and the curve where the circle is located represents the data obtained from the training data set through the least squares method.

[0093] It can be seen that when the training error is 0.031817, the generalization error estimate is 2.406155; when the training error is 0.027910, the generalization error estimate is 1.220573; when the training error is 0.027999, the generalization error estimate is 0.422798; when the training error is 0.032610, the generalization error estimate is 0.043315; when the training error is 0.034308, the generalization error estimate is 0.407059; when the training error is 0.028333, the generalization error estimate is 1.237353.

[0094] It can be seen that for the sub-model with the smallest training error of 0.0027910, the curve where the circle is located and the curve where the triangle is located are very different, so the sub-model fits poorly. However, for the sub-model with a training error of 0.032610, the curve where the circle is located and the curve where the triangle is located fit better.

[0095] Therefore, this sub-model can be used as the initial value of the medical model, and the communication requirements in the next cloud update will be greatly reduced. In this example, the cloud update uses the gradient descent method. It can be seen that the final cloud model after fusion is highly consistent with the real model.

[0096] Based on the same inventive concept, Figure 7 is a block diagram of a device for improving data processing accuracy according to an exemplary embodiment. Figure 7 As shown, the device 500 includes: an acquisition module 510 and an execution module 520.

[0097] Wherein, the acquisition module 510 is configured to determine target data information of a target user;

[0098] The execution module 520 is configured to obtain the data feature value of the target user according to the target data information of the target user based on the called cloud neural network model, wherein the cloud neural network model is trained in the following manner: obtaining a plurality of sub-models, each of which is trained based on the original data information of the seed user as a training sample, using the plurality of sub-models as target training samples of the cloud neural network model, and training to obtain the cloud neural network model;

[0099] The cloud-based neural network model is used to use the target data information as input to determine the target features of the target user, and output the target features as output.

[0100] In this way, the accuracy of data processing can be improved, and the accuracy of user classification can be improved. In addition, an accurate neural network model can be trained without leaving the data source, which can improve the accuracy of user classification. In addition, multiple sub-models can be fused into a new model in the cloud, and its structure is consistent with the model structure of the sub-model, without changing the model architecture corresponding to each sub-model, which reduces the amount of calculation and saves costs.

[0101] Preferably, the multiple sub-models are used as target training samples of the cloud neural network model to train the cloud neural network model, specifically:

[0102] Calculate the complexity of each sub-model with respect to the corresponding training sample to obtain the generalization error corresponding to the sub-model;

[0103] Using the sub-model with a generalization error less than or equal to a preset threshold as the target training sample for training the cloud neural network model;

[0104] The cloud-based neural network model is trained based on the target training samples.

[0105] Preferably, the calculation of the complexity of each sub-model with respect to the corresponding training sample to obtain the generalization error corresponding to the sub-model is specifically as follows:

[0106]

[0107] Among them, L D (h) is the generalization error, L S (h) is the training error, is the Rademacher complexity, is the loss function, is the hypothesis class, S is the number of samples in the training set, and L is the loss function The constant for Lipschitz continuity of the prediction term, r>0 is the radius of the local Rademacher complexity, which is used to constrain the hypothesis class To obtain a tighter upper bound, λ j , j = 1...m is a symmetric square matrix Eigenvalues ​​in non-ascending order, It represents the value of the i-th sample propagated to the layer immediately adjacent to the output layer, σ is the confidence level, and 0<σ<1. Preferably, the confidence level σ is 0.95.

[0108] Preferably, the cloud-based neural network model training further includes:

[0109] Requesting the sub-model for the amount of data of the training samples of the sub-model and the gradient of the training error in the model parameters;

[0110] Calculating the gradient direction of the cloud neural network model according to the data volume of the training samples and the gradient of the training error in the model parameters;

[0111] Based on the gradient descent method, the cloud neural network model is updated according to the gradient direction;

[0112] Determine whether the updated cloud-based neural network model converges, and if the cloud-based neural network model does not converge, recalculate the gradient direction of the cloud-based neural network model.

[0113] Preferably, the gradient direction of the cloud neural network model is calculated as follows:

[0114]

[0115] in, is the gradient direction, is the sample size of the training set.

[0116] Based on the same inventive concept, the present disclosure also provides an electronic device, including:

[0117] A memory storing programmable control instructions;

[0118] A processor is used to execute the programmable control instructions in the memory to implement the steps of any method described in the first aspect.

[0119] Based on the same inventive concept, an embodiment of the present disclosure provides a computer-readable storage medium having computer program instructions stored thereon, which, when executed by a processor, implement the steps of any one of the methods described in the first aspect.

[0120] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the present disclosure. This application is intended to cover any variations, uses or adaptations of the present disclosure, which follow the general principles of the present disclosure and include common knowledge or customary technical means in the art that are not disclosed in the present disclosure. The specification and examples are to be considered as exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0121] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A method for improving data processing accuracy, characterized in that: The method comprises: Determine target data information of target users; Based on the called cloud neural network model, the target features of the target user are obtained according to the target data information of the target user, wherein the cloud neural network model is trained in the following manner: a plurality of sub-models are obtained, each of which is obtained by training based on the original data information of the seed user as a training sample, and the plurality of sub-models are used as target training samples of the cloud neural network model to obtain the cloud neural network model through training; The cloud neural network model is used to use the target data information as input to determine the target features of the target user, and output the target features as output; Wherein, the neural network model is a medical model, firstly, m sub-models are trained locally based on the original medical records and original clinical data of the patient to obtain multiple sub-models, then the m sub-models obtained by training are obtained, and the complexity of each sub-model with respect to the training data is calculated to obtain the generalization error corresponding to the sub-model, so as to evaluate the sub-model, and then determine the target sub-model as the initial value, and upload each evaluation result to the cloud so that the cloud can select the target sub-model as the initial value, and, according to the evaluation result, select the sub-model as the initial value, and train to obtain the medical model, judge whether the obtained medical model needs to be updated, and when it is determined that the medical model does not need to be updated, obtain the fused model, and when it is determined that the medical model needs to be updated, request the data required by the medical model from the local, which is non-training data, and receive k feedbacks, and then update the medical model, wherein 0 <k<m+1; The step of using the multiple sub-models as target training samples of the cloud-based neural network model to train the cloud-based neural network model includes: Calculate the complexity of each sub-model with respect to the corresponding training sample to obtain the generalization error corresponding to the sub-model; Using the sub-model with a generalization error less than or equal to a preset threshold as the target training sample for training the cloud neural network model; According to the target training sample, the cloud neural network model is trained; The calculation of the complexity of each sub-model with respect to the corresponding training sample to obtain the generalization error corresponding to the sub-model is specifically as follows: Among them, L D (h) is the generalization error, L S (h) is the training error, is the Rademacher complexity, l is the loss function, is the hypothesis class, S is the number of samples in the training set, and L is the constant of loss function l with respect to the prediction term Lipschitz continuity, r>0 is the radius of local Rademacher complexity, which is used to constrain the hypothesis class To obtain a tighter upper bound, λ j A symmetric square matrix Eigenvalues ​​in non-ascending order, j = 1...m, represents the value of the i-th sample propagated to the layer immediately adjacent to the output layer, δ is the confidence level, and 0<δ<1, the confidence level δ is 0.95; The cloud-based neural network model training also includes: Requesting the sub-model for the amount of data of the training samples of the sub-model and the gradient of the training error in the model parameters; Calculating the gradient direction of the cloud neural network model according to the data volume of the training samples and the gradient of the training error in the model parameters; Based on the gradient descent method, the cloud neural network model is updated according to the gradient direction; Determine whether the updated cloud neural network model converges, and if the cloud neural network model does not converge, recalculate the gradient direction of the cloud neural network model; The calculation of the gradient direction of the cloud neural network model is specifically as follows: in, is the gradient direction, is the sample size of the training set.

2. A device for improving data processing accuracy, characterized in that: The device comprises: An acquisition module is configured to determine target data information of a target user; The execution module is configured to obtain the data features of the target user according to the target data information of the target user based on the called cloud neural network model, wherein the cloud neural network model is trained in the following manner: obtaining a plurality of sub-models, each of which is trained based on the original data information of the seed user as a training sample, using the plurality of sub-models as target training samples of the cloud neural network model, and training to obtain the cloud neural network model; The cloud neural network model is used to use the target data information as input to determine the target features of the target user, and output the target features as output; Wherein, the neural network model is a medical model, firstly, m sub-models are trained locally based on the original medical records and original clinical data of the patient to obtain multiple sub-models, then the m sub-models obtained by training are obtained, and the complexity of each sub-model with respect to the training data is calculated to obtain the generalization error corresponding to the sub-model, so as to evaluate the sub-model, and then determine the target sub-model as the initial value, and upload each evaluation result to the cloud so that the cloud can select the target sub-model as the initial value, and, according to the evaluation result, select the sub-model as the initial value, and train to obtain the medical model, judge whether the obtained medical model needs to be updated, and when it is determined that the medical model does not need to be updated, obtain the fused model, and when it is determined that the medical model needs to be updated, request the data required by the medical model from the local, which is non-training data, and receive k feedbacks, and then update the medical model, wherein 0 <k<m+1; The multiple sub-models are used as target training samples of the cloud neural network model to train the cloud neural network model, specifically: Calculate the complexity of each sub-model with respect to the corresponding training sample to obtain the generalization error corresponding to the sub-model; Using the sub-model with a generalization error less than or equal to a preset threshold as the target training sample for training the cloud neural network model; According to the target training sample, the cloud neural network model is trained; The calculation of the complexity of each sub-model with respect to the corresponding training sample to obtain the generalization error corresponding to the sub-model is specifically as follows: Among them, L D (h) is the generalization error, L S (h) is the training error, is the Rademacher complexity, l is the loss function, is the hypothesis class, S is the number of samples in the training set, and L is the constant of loss function l with respect to the prediction term Lipschitz continuity, r>0 is the radius of local Rademacher complexity, which is used to constrain the hypothesis class To obtain a tighter upper bound, λ j A symmetric square matrix Eigenvalues ​​in non-ascending order, represents the value of the i-th sample propagated to the layer immediately adjacent to the output layer, δ is the confidence level, and 0<δ<1, the confidence level δ is 0.95; The cloud-based neural network model training also includes: Requesting the sub-model for the amount of data of the training samples of the sub-model and the gradient of the training error in the model parameters; Calculating the gradient direction of the cloud neural network model according to the data volume of the training samples and the gradient of the training error in the model parameters; Based on the gradient descent method, the cloud neural network model is updated according to the gradient direction; Determine whether the updated cloud neural network model converges, and if the cloud neural network model does not converge, recalculate the gradient direction of the cloud neural network model; The calculation of the gradient direction of the cloud neural network model is specifically as follows: in, is the gradient direction, is the sample size of the training set.

3. An electronic device, characterized in that: include: A memory storing programmable control instructions; A processor, configured to execute the programmable control instructions in the memory to implement the steps of the method of claim 1.

4. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the program instructions are executed by a processor, the steps of the method according to claim 1 are implemented.

Images