Model optimization method, device and system

Abstract

The embodiment of the invention provides a model optimization method, device and system.The model optimization method is applied to a client side and comprises the steps that global model parameters issued by a server are obtained, a service prediction model is constructed according to the global model parameters, and the service prediction model is used for processing a multimedia data processing task; determining reference data representation and sample distribution information corresponding to the business prediction model, and selecting local sample data matched with the sample distribution information from a local sample set to form an initial sample set; determining sample distribution difference information according to the sample data representation corresponding to the initial sample set and the reference data representation, and calculating a sample distribution constraint condition according to the sample distribution difference information; adjusting the initial sample set according to the sample distribution constraint condition to obtain a target sample set; and optimizing the service prediction model by using the target sample set, and feeding back model parameters of the optimized service prediction model to the server.

Description

Technical Field

[0001] The embodiments of this specification relate to the field of machine learning technology, and in particular to model optimization methods, devices and systems. Background Art

[0002] With the development of computer and Internet technology, federated learning has been applied in more and more scenarios. Federated learning can well solve the problem of local data isolation, but it requires a new machine learning paradigm to improve the utilization of multi-party knowledge and the performance of the model through joint modeling. Federated learning focuses on modeling distributed data in order to promote the utilization of information of multi-party data by the model without leaking privacy and improve the modeling ability. Non-IID (Non-Independent and identically distributed, Non-IID) and distributed data allow the data modeling of each party to have personalized performance, which will have a cascading effect in joint modeling. In addition, the data is distributed in multiple non-independent and identically distributed data sources, resulting in the model mapping different data to a limited representation space, and the high-dimensional representation vectors become highly correlated and concentrated in a low-dimensional subspace, and cannot fully utilize the entire representation space, which leads to information loss and hinders the model from capturing the full complexity of the data. In addition, in federated learning, multiple local models and global models are modeled collaboratively, which further causes the problem of representation collapse coupling, resulting in a local to global cascade effect, which not only affects the local model itself, but also affects the global model and other local models participating in the federated learning process, resulting in poor performance and hindering the convergence of the learning process. Representation collapse indicates that the diversity and richness of learned representations are reduced, while representation collapse coupling highlights the inter-correlation of representation collapse between different parts of the model or different models in a federated learning setting. Therefore, an effective solution is urgently needed to solve the above problems. Summary of the invention

[0003] In view of this, an embodiment of this specification provides a model optimization method. One or more embodiments of this specification also relate to a model optimization device, a model optimization system, a computing device, a computer-readable storage medium and a computer program product to solve the technical defects existing in the prior art.

[0004] According to a first aspect of an embodiment of this specification, a model optimization method is provided, which is applied to a client and includes: Acquire global model parameters sent by the server, and construct a business prediction model according to the global model parameters, wherein the business prediction model is used to process multimedia data processing tasks; Determine the benchmark data representation and sample distribution information corresponding to the business prediction model, and select local sample data matching the sample distribution information from the local sample set to form an initial sample set; Determine sample distribution difference information according to the sample data representation corresponding to the initial sample set and the benchmark data representation, and calculate sample distribution constraint conditions according to the sample distribution difference information; Adjusting the initial sample set according to the sample distribution constraint to obtain a target sample set; The business prediction model is optimized using the target sample set, and the model parameters of the optimized business prediction model are fed back to the server.

[0005] Optionally, the acquiring global model parameters sent by the server and constructing a business prediction model according to the global model parameters includes: Acquire global model parameters sent by the server, wherein the global model parameters are obtained by the server aggregating local model parameters submitted by multiple clients; Loading an initial business prediction model, and updating the initial business prediction model using the global model parameters to obtain a business prediction model; The multimedia data processing tasks include text processing tasks, image processing tasks, video processing tasks or audio processing tasks.

[0006] Optionally, before the step of selecting local sample data matching the sample distribution information in the local sample set to form an initial sample set is performed, the method further includes: Determine spherical Gaussian distribution information preset for the business prediction model, and use the spherical Gaussian distribution information as sample distribution information; The step of selecting local sample data matching the sample distribution information from the local sample set to form an initial sample set includes: Determine a local sample set corresponding to the business prediction model, and select a plurality of local sample data from the local sample set according to the spherical Gaussian distribution information; An initial sample set is constructed based on the multiple local sample data, wherein distribution information corresponding to the initial sample set is the same as the spherical Gaussian distribution information.

[0007] Optionally, determining sample distribution difference information according to the sample data representation corresponding to the initial sample set and the benchmark data representation, and calculating sample distribution constraint conditions according to the sample distribution difference information, includes: Determining a sample data representation corresponding to the initial sample set; Determining sample distribution difference information by comparing the sample data representation and the benchmark data representation; The sample distribution constraint divergence is calculated according to a preset optimization strategy and the sample distribution difference information, and the sample distribution constraint divergence is used as a sample distribution constraint condition.

[0008] Optionally, optimizing the business prediction model by using the target sample set and feeding back model parameters of the optimized business prediction model to the server includes: Selecting target sample data from the target sample set, and optimizing the business prediction model using the target sample data; Check whether the optimized business prediction model meets the local optimization stop condition; If not, selecting candidate sample data from the target sample set, using the candidate sample data as the target sample data, and performing a step of optimizing the business prediction model using the target sample data; If so, the model parameters of the optimized business prediction model are fed back to the server.

[0009] According to a second aspect of an embodiment of this specification, a model optimization device is provided, which is applied to a client and includes: An acquisition module is configured to acquire global model parameters sent by a server, and construct a business prediction model according to the global model parameters, wherein the business prediction model is used to process multimedia data processing tasks; A determination module is configured to determine the benchmark data representation and sample distribution information corresponding to the business prediction model, and select local sample data matching the sample distribution information from the local sample set to form an initial sample set; A calculation module, configured to determine sample distribution difference information according to the sample data representation corresponding to the initial sample set and the benchmark data representation, and calculate sample distribution constraint conditions according to the sample distribution difference information; An adjustment module is configured to adjust the initial sample set according to the sample distribution constraint to obtain a target sample set; The optimization module is configured to optimize the business prediction model using the target sample set and feed back the model parameters of the optimized business prediction model to the server.

[0010] According to a third aspect of an embodiment of this specification, a model optimization system is provided, the model optimization system comprising a server and a plurality of clients, including: A target client among the multiple clients is used to receive the global model parameters sent by the server, and to construct a business prediction model according to the global model parameters, wherein the business prediction model is used to process multimedia data processing tasks; determine the benchmark data representation and sample distribution information corresponding to the business prediction model, and select local sample data matching the sample distribution information in the local sample set to form an initial sample set; determine sample distribution difference information according to the sample data representation corresponding to the initial sample set and the benchmark data representation, and calculate sample distribution constraints according to the sample distribution difference information; adjust the initial sample set according to the sample distribution constraints to obtain a target sample set; optimize the business prediction model using the target sample set, and feed back the model parameters of the optimized business prediction model to the server; The server is used to receive model parameters uploaded by each client, build a target global business prediction model that meets the training stop condition based on the model parameters uploaded by each client, and send the target global model parameters corresponding to the target global business prediction model to each client.

[0011] According to a fourth aspect of an embodiment of this specification, a computing device is provided, including: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the above-mentioned model optimization method are implemented.

[0012] According to a fifth aspect of the embodiments of this specification, a computer-readable storage medium is provided, which stores computer-executable instructions, and when the instructions are executed by a processor, the steps of the above-mentioned model optimization method are implemented.

[0013] According to a sixth aspect of the embodiments of this specification, a computer program product is provided, comprising a computer program or instructions, which implement the steps of the above-mentioned model optimization method when executed by a processor.

[0014] In order to avoid the collapse problem of sample distribution, the model optimization method provided in this embodiment can perform a model optimization stage on the client side, first build a business prediction model according to the global model parameters sent by the server, and the business prediction model is used to process multimedia data processing tasks; on this basis, the benchmark data representation and sample distribution information corresponding to the business prediction model can be determined, and local sample data matching the sample distribution information can be selected from the local sample set to form an initial sample set; the sampled initial sample data can be consistent with the sample distribution information, thereby avoiding the collapse of the representation, thereby effectively improving the performance of the client side, and making full use of the complete representation space to obtain rich and diverse distinguishable representation information. On this basis, considering that the initial sample data obtained by sampling according to the sample distribution information has the problem of category separation, that is, samples of the same category are over-dispersed and lose similarity; therefore, the sample distribution difference information can be determined according to the sample data representation and the benchmark data representation corresponding to the initial sample set, and the sample distribution constraint conditions can be calculated according to the sample distribution difference information; at this time, the initial sample set is adjusted by introducing sample distribution constraints, and then the target sample set can be obtained; in the sample sampling stage, a sample set that eliminates the above problems can be obtained, and finally the target sample set can be used to optimize the business prediction model, and the model parameters of the optimized business prediction model are fed back to the server. The server cooperates with multiple clients to complete federated learning, which can avoid the cascading effect caused by the collapse of the representation, thereby realizing multi-party joint training to obtain a business prediction model that meets the usage needs and has stronger generalization ability for use in multimedia data processing tasks to improve the prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a schematic diagram of a model optimization method provided by an embodiment of this specification; Figure 2 is a flow chart of a model optimization method provided by an embodiment of this specification; Figure 3 is a structural schematic diagram of a model optimization device provided by an embodiment of this specification; Figure 4 is a structural schematic diagram of a model optimization system provided by an embodiment of this specification; Figure 5 It is a structural block diagram of a computing device provided by an embodiment of this specification. DETAILED DESCRIPTION

[0016] Many specific details are described in the following description to facilitate a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of this specification, so this specification is not limited to the specific implementation disclosed below.

[0017] The terms used in one or more embodiments of this specification are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of this specification. The singular forms of "a", "said" and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0018] It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of this specification, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0019] In addition, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0020] First, the terms involved in one or more embodiments of this specification are explained.

[0021] Cascade effect is an interdisciplinary concept, which refers to the effect of an initial action or event affecting a system, which in turn triggers a series of continuous and interrelated events or processes. This effect has a wide range of applications and manifestations in many fields, including but not limited to aviation, medicine, ecology, computer science, physics, and disaster management.

[0022] Representation collapse means that during the training process, the internal representation of the model (i.e., hidden layer output or feature vector) gradually loses the ability to distinguish differences in input data, making the representation of all input data similar or consistent.

[0023] In this specification, a model optimization method is provided. This specification also involves a model optimization device, a model optimization system, a computing device, a computer-readable storage medium and a computer program product, which are described in detail one by one in the following embodiments.

[0024] See also Figure 1 As shown in the schematic diagram, the model optimization method provided in this embodiment, in order to avoid the collapse problem of sample distribution, can perform model optimization in the client, first build a business prediction model according to the global model parameters sent by the server, and the business prediction model is used to process multimedia data processing tasks; on this basis, the benchmark data representation and sample distribution information corresponding to the business prediction model can be determined, and local sample data matching the sample distribution information can be selected from the local sample set to form an initial sample set; the sampled initial sample data can be consistent with the sample distribution information, thereby avoiding the collapse of the representation, thereby effectively improving the performance of the client, and making full use of the complete representation space to obtain rich and diverse distinguishable representation information. On this basis, considering that the initial sample data obtained by sampling according to the sample distribution information has the problem of category separation, that is, samples of the same category are over-dispersed and lose similarity; therefore, the sample distribution difference information can be determined according to the sample data representation and the benchmark data representation corresponding to the initial sample set, and the sample distribution constraint conditions can be calculated according to the sample distribution difference information; at this time, the initial sample set is adjusted by introducing sample distribution constraints, and then the target sample set can be obtained; in the sample sampling stage, a sample set that eliminates the above problems can be obtained, and finally the target sample set can be used to optimize the business prediction model, and the model parameters of the optimized business prediction model are fed back to the server. The server cooperates with multiple clients to complete federated learning, which can avoid the cascading effect caused by the collapse of the representation, thereby realizing multi-party joint training to obtain a business prediction model that meets the usage needs and has stronger generalization ability for use in multimedia data processing tasks to improve the prediction effect.

[0025] See also Figure 2 , Figure 2 A flowchart of a model optimization method provided according to an embodiment of the present specification is shown, which is applied to a client and specifically includes the following steps.

[0026] Step S202, obtaining global model parameters sent by the server, and constructing a business prediction model according to the global model parameters, wherein the business prediction model is used to process multimedia data processing tasks.

[0027] The model optimization method provided in this embodiment is applied to any client participating in federated learning in a federated learning scenario, and the execution logic of any client in the federated learning scenario is the same as the model optimization method provided in this embodiment. The business prediction model jointly trained by multiple parties in a federated learning scenario can be a model for processing multimedia data processing tasks, such as text processing tasks, used for question answering, translation, abstract extraction, etc.; it can also be an image processing task for image enhancement, color adjustment, image generation, etc. It can also be an audio processing task for audio optimization, audio generation, audio noise elimination, etc. In actual applications, the application scenarios and prediction capabilities of the models jointly trained by multiple parties can be set according to actual needs. This embodiment does not make any restrictions here.

[0028] Specifically, the global model parameters refer to the model parameters obtained by aggregating the model parameters corresponding to the local business prediction models of multiple clients by the server in the federated learning scenario, so as to achieve the aggregation of the knowledge of each client to ensure the completion of the federated training model without leaking data. Correspondingly, the business prediction model is the prediction model to be trained with the knowledge capabilities of other clients obtained by adjusting the local model of the client in combination with the global model parameters. Correspondingly, the multimedia data processing tasks specifically refer to the tasks that need to be performed in the scenario where the business prediction model is deployed, including but not limited to text processing tasks, image processing tasks, video processing tasks or audio processing tasks.

[0029] This embodiment uses a multi-party jointly trained business prediction model for a text processing task as an example to illustrate the model optimization method, which is used to provide an answer text after inputting a question text. The model optimization methods in other scenarios can refer to the same or corresponding descriptions in this embodiment, and this embodiment will not be described in detail here.

[0030] In practical applications, in federated learning, representation collapse not only reduces the performance of local clients, but also intricately affects the representation of global and local models. This embodiment provides a model optimization method that can improve the model's utilization efficiency of the representation space while ensuring that privacy data is not leaked, increase the diversity and richness of the model's representation, and promote the generalization ability of the model. Compared with the existing federated model, the model optimization method provided in this embodiment can have no constraints on the training objectives and modeling models of federated learning, and directly perform representation distribution maximization constraints in the representation space of the model, thereby increasing the utilization of the representation space without affecting the similarity of similar samples, and solving the problem of representation collapse. Subsequently, it prevents the representation collapse from further coupling with other models, so that the representation in the entire federated modeling environment is effectively utilized, and the model performance and generalization effect are improved.

[0031] That is to say, in order to avoid the collapse problem of sample distribution, during the model optimization stage on the client side, a business prediction model can be built according to the global model parameters sent by the server, and the business prediction model is used to process multimedia data processing tasks; on this basis, the benchmark data representation and sample distribution information corresponding to the business prediction model can be determined, and local sample data matching the sample distribution information can be selected from the local sample set to form an initial sample set; the initial sample data sampled can conform to the sample distribution information, thereby avoiding the collapse of the representation, thereby effectively improving the performance of the client, and making full use of the complete representation space to obtain rich and diverse distinguishable representation information. On this basis, considering that the initial sample data obtained by sampling according to the sample distribution information has the problem of category separation, that is, samples of the same category are over-dispersed and lose similarity; therefore, the sample distribution difference information can be determined according to the sample data representation and the benchmark data representation corresponding to the initial sample set, and the sample distribution constraint conditions can be calculated according to the sample distribution difference information; at this time, the initial sample set is adjusted by introducing sample distribution constraints, and then the target sample set can be obtained; in the sample sampling stage, a sample set that eliminates the above problems can be obtained, and finally the target sample set can be used to optimize the business prediction model, and the model parameters of the optimized business prediction model are fed back to the server. The server cooperates with multiple clients to complete federated learning, which can avoid the cascading effect caused by the collapse of the representation, thereby realizing multi-party joint training to obtain a business prediction model that meets the usage needs and has stronger generalization ability for use in multimedia data processing tasks to improve the prediction effect.

[0032] Furthermore, when the business prediction model is constructed according to the global model parameters sent by the server, it is actually an adjustment of the local model based on the global model parameters. In this embodiment, the specific implementation method is as follows: Obtain global model parameters issued by the server, wherein the global model parameters are obtained by aggregating local model parameters submitted by multiple clients by the server; load an initial business prediction model, and use the global model parameters to update the initial business prediction model to obtain a business prediction model; wherein the multimedia data processing task includes a text processing task, an image processing task, a video processing task or an audio processing task.

[0033] Specifically, the initial business prediction model specifically refers to the business prediction model deployed locally on the client. Based on this, when the global model parameters issued by the server aggregating the local model parameters submitted by multiple clients are obtained, the initial business prediction model can be loaded first, and then the global model parameters can be used to update the initial business prediction model, so that the initial business prediction model can be adjusted to the business prediction model, and then the business prediction model can be trained using local samples, and then aggregated by the server to obtain a global model that meets the usage requirements. Among them, multimedia data processing tasks include text processing tasks, image processing tasks, video processing tasks or audio processing tasks.

[0034] In the specific implementation, for the business prediction model corresponding to each client, the training objective not only needs to maintain the invariance of the same samples with different augmentations, but also needs to expand the uniform distribution property of a given batch of B representations, that is, for the representation information after two types of augmented samples, and , we train the self-supervised model to minimize the overall goal of formula (1):

[0035] in, is an alignment term, which uses the bi-norm form to constrain the two types of augmented representations to be consistent. is a uniform distribution constraint, which is used to expand the space of sample representation. is a hyperparameter that balances these two terms. The alignment term makes data samples of the same category clustered, while other samples are separable, that is, The uniform distribution constraint enhances the consistency of representation and is the key to reducing representation collapse. In order to reduce the reliance on prior knowledge of client data, the local sample representation can be normalized into a random distribution with a high entropy value. Then the business prediction model can be trained according to this distribution.

[0036] For example, when Party A, Party B, and Party C jointly train a text processing model with multi-domain question-answering capabilities, it is determined that Party A holds text data in the computer field, Party B holds text data in the transaction field, and Party C holds text data in the education field. On this basis, Party A, Party B, and Party C will first train a local model corresponding to their own field in combination with the locally stored text data, and then aggregate the model parameters corresponding to the local model to the server. The server aggregates the model parameters submitted by each participant according to the preset aggregation rules. After obtaining the global model parameters, the global model parameters will be sent to Party A, Party B, and Party C, so that Party A, Party B, and Party C can update the local model based on the global model parameters, so that the local model can learn the knowledge and capabilities of other participants. On this basis, the updated local model will continue to be trained using local data to train a text processing model with question-answering capabilities in multiple fields.

[0037] In summary, by adjusting the initial business prediction model using the global model parameters sent by the server, the corresponding knowledge of other participants can be integrated, and then a business prediction model that meets usage needs can be trained under the premise of ensuring data security in a multi-party joint scenario.

[0038] Step S204, determining the benchmark data representation and sample distribution information corresponding to the business prediction model, and selecting local sample data matching the sample distribution information from the local sample set to form an initial sample set.

[0039] Specifically, after obtaining the business prediction model as described above, when using local sample data to train the business prediction model, in order to avoid the problem of representation collapse, the benchmark data representation and sample distribution information corresponding to the business prediction model can be determined first, and on this basis, local sample data can be selected from the local sample set according to the sample distribution information to form an initial sample set. The initial sample data of random sampling can match the sample distribution information, thereby increasing the utilization rate of the representation space and solving the problem of representation collapse.

[0040] Among them, the benchmark data representation specifically refers to the normalized data representation pre-set by the business prediction model. Although it can achieve sample processing in accordance with the normalized standard, it will also bring about the problem of representation collapse caused by unreasonable sample distribution. Therefore, when sampling samples, the preset sample distribution information can be determined, and then sampling can be performed according to the sample distribution information. Correspondingly, the sample distribution information is the information that ensures the reasonable sample distribution after the preset sample sampling for the business prediction model. Sample sampling according to the sample distribution information can ensure that the samples will not be over-dispersed or over-aggregated.

[0041] Furthermore, in order to avoid characterization collapse, a spherical Gaussian distribution can be used to collect samples. In this embodiment, the specific implementation is as follows: Determine the spherical Gaussian distribution information preset for the business prediction model, and use the spherical Gaussian distribution information as sample distribution information; wherein, the local sample data matching the sample distribution information is selected in the local sample set to form an initial sample set, including: determining the local sample set corresponding to the business prediction model, and selecting multiple local sample data in the local sample set according to the spherical Gaussian distribution information; constructing an initial sample set based on the multiple local sample data, wherein the distribution information corresponding to the initial sample set is the same as the spherical Gaussian distribution information.

[0042] Based on this, in order to alleviate the problem of representation collapse, we can first determine the spherical Gaussian distribution information preset for the business prediction model, and use the spherical Gaussian distribution information as the sample distribution information; on this basis, in the sampling stage, we can determine the local sample set corresponding to the business prediction model, and select multiple local sample data from the local sample set according to the spherical Gaussian distribution information; so as to build an initial sample set based on multiple local sample data, and the distribution information corresponding to the initial sample set is the same as the spherical Gaussian distribution information, thereby ensuring that the representation collapse problem is solved without leaking data.

[0043] Specifically, we can choose samples that follow a spherical Gaussian distribution, i.e., , so that , as a priori. In this way, the representation collapse is mitigated and information leakage is avoided in the federated representation process. After that, each participant’s corresponding local model has a set of standardized data representations , can consistently follow a spherical Gaussian distribution Sampling a set of random samples, we can then measure the distribution difference between the two by comparing the same spherical Gaussian distribution, which can be achieved by the following formula (2):

[0044] Therefore, the uniform distribution constraint Even dispersion avoids representation collapse without excluding instance-based contrastive learning.

[0045] Continuing with the above example, for any one of Party A, Party B and Party C, when sampling local text data to optimize the updated local model, in order to avoid data leakage and alleviate the representation collapse problem, they can first determine the normalized data representation corresponding to the local model, and then perform random sampling according to the spherical Gaussian distribution, and then obtain an initial text data set composed of initial text data matching the spherical Gaussian distribution, so that the local model can be trained based on the set in the future to enable it to have text processing capabilities.

[0046] In summary, in order to alleviate the problem of representation collapse, spherical Gaussian distribution can be used to construct the initial sample data set, so that the data can fully utilize the representation space and avoid data leakage.

[0047] Step S206 , determining sample distribution difference information according to the sample data representation corresponding to the initial sample set and the benchmark data representation, and calculating sample distribution constraint conditions according to the sample distribution difference information.

[0048] Step S208: adjusting the initial sample set according to the sample distribution constraint to obtain a target sample set.

[0049] Specifically, after obtaining the sample distribution information and benchmark data representation of the corresponding model, since the client data is Non-IID, when the sample representation is strictly constrained to be close to a random instance, the category separation will be destroyed, that is, the samples of the same category lose similarity because they are too scattered. Therefore, in order to ensure that the samples used in the training model are more effective, the sample distribution difference information can be determined based on the sample data representation and benchmark data representation corresponding to the initial sample set. On this basis, the sample distribution constraint conditions can be calculated based on the sample distribution difference information, and then the initial sample set can be adjusted according to the sample distribution constraint conditions to generate the target sample set according to the adjustment results.

[0050] Specifically, the sample data representation specifically refers to the data representation corresponding to the initial sample data in the initial sample set. Correspondingly, the sample distribution difference information specifically refers to the difference information obtained by comparing the sample data representation corresponding to the initial sample set obtained after sampling according to the sample distribution information with the normalized benchmark data representation preset by the model, so as to reflect the separation of the current sample data, and then the sample distribution constraint conditions can be constructed in combination with the sample distribution difference information, so that after adjusting the initial sample set, the target sample set obtained after the adjustment can satisfy the requirements of reducing the representation collapse and not causing the samples to be too dispersed and lose similarity. Correspondingly, the target sample set specifically refers to the sample data set obtained after optimizing the initial sample set.

[0051] Furthermore, when calculating the sample distribution constraint condition, the sample distribution constraint triplet can be performed in combination with the sample distribution difference information. In this embodiment, the specific implementation method is as follows: Determine the sample data representation corresponding to the initial sample set; determine the sample distribution difference information by comparing the sample data representation and the benchmark data representation; calculate the sample distribution constraint divergence according to a preset optimization strategy and the sample distribution difference information, and use the sample distribution constraint divergence as a sample distribution constraint condition.

[0052] Specifically, the sample distribution constraint divergence refers to the UOT (Unbalanced Optimal Transport) divergence. Based on this, in order to reduce the problem of representation collapse and prevent the sample data from being too dispersed and losing similarity, the sample data representation corresponding to the initial sample set can be determined first; then the sample distribution difference information can be determined by comparing the sample data representation with the benchmark data representation; on this basis, the sample distribution constraint divergence can be calculated according to the preset optimization strategy and the sample distribution difference information, and the sample distribution constraint divergence can be used as the sample distribution constraint condition for subsequent adjustment of the initial sample set.

[0053] In practical applications, when solving the problem of over-dispersed samples, data samples can be matched with random Gaussian samples with arbitrary or proportional mass, that is, the sampling pairs with lower uncertainty are kept unmatched. Therefore, unbalanced optimal transport (UOT) is one of the effective solutions. UOT calculates the transmission mapping between two sample masses with different distributions under soft margin constraints, for example, the prediction edge and the ground truth edge. -Normalization. Given data and Gaussian distribution with marginal constraints and , then we can formulate a search for the pairing matrix with the minimum distribution difference The UOT problem is as follows:

[0054] Among them, the cost matrix ,and ) is an indicator for row (column) Kronecker multiplication, where Represents the identity matrix.

[0055] When optimizing, you can and , then equation (3) can be rewritten into a positive definite quadratic form, as follows:

[0056] in is a constant, that is, there exists the following relationship in formula (5):

[0057] Next, we can optimize by steepest gradient descent , as follows formula (6):

[0058] Among them, the optimal learning rate is defined as follows:

[0059] Finally, by taking the optimal Bringing back formula (3), we can obtain a uniform UOT divergence. FUR can normalize data samples close to a spherical Gaussian distribution by minimizing the UOT divergence. It should be noted that the spherical Gaussian distribution maximizes its entropy and uniformly distributes its samples. The mapped data representation enjoys the above-mentioned good properties of the spherical Gaussian distribution and further alleviates the representation collapse coupling.

[0060] Continuing with the above example, after obtaining the initial text data set collected by the participants according to the spherical Gaussian distribution, in order to avoid the loss of similarity of samples, constraints can be added to the initial text data set according to the above processing method, so that the text data can be optimized and constrained in the representation space, thereby obtaining the target text data set corresponding to the participants, so as to facilitate the subsequent text processing model training.

[0061] In summary, by adjusting samples in the above manner, it is possible to achieve alignment of the representations of all local models with the spherical Gaussian distribution without relying on local data distribution, reducing the problem of high correlation of representations in low-dimensional space, and by the above manner, any representation space can be optimized and constrained, thereby adapting to different models and frameworks.

[0062] Step S210: Optimize the business prediction model using the target sample set, and feed back the model parameters of the optimized business prediction model to the server.

[0063] Specifically, after obtaining the optimized target sample set as mentioned above, the target sample set can be used to optimize the business prediction model, and the model parameters of the optimized business prediction model can be fed back to the server, so that the server can aggregate the model parameters corresponding to multiple clients and then perform federal training, and so on, until a model that meets the training stop conditions is obtained and published to each client, and each client can be supported to use it, thereby ensuring that the model has stronger prediction capabilities without leaking data.

[0064] Furthermore, in the model optimization stage, it is necessary to continuously optimize the business prediction model in combination with the target sample data contained in the target sample set. In this embodiment, the specific implementation method is as follows: Select target sample data from the target sample set, and use the target sample data to optimize the business prediction model; detect whether the optimized business prediction model meets the local optimization stop condition; if not, select candidate sample data from the target sample set, use the candidate sample data as the target sample data, and execute the step of optimizing the business prediction model using the target sample data; if so, feed back the model parameters of the optimized business prediction model to the server.

[0065] Based on this, when the client uses the target sample data to train the business prediction model, the target sample data can be selected from the target sample set, and the business prediction model can be optimized using the target sample data; at this time, it can be detected whether the optimized business prediction model meets the local optimization stop condition; if not, it means that training needs to continue, so candidate sample data can be selected from the target sample set, and the candidate sample data can be used as the target sample data, and the step of optimizing the business prediction model using the target sample data is executed; if so, it means that the model can be fed back to the server, so the model parameters of the optimized business prediction model can be fed back to the server for model aggregation in the next cycle.

[0066] Continuing with the above example, after obtaining the target text data set, the text data contained in the target text data set can be used to train the text processing model. When the preset stop condition is met after training, the model parameters corresponding to the model that meets the condition can be sent to the server, and the server aggregates the model parameters corresponding to Party A, Party B and Party C respectively, and then sends them to each client for a new round of model training. After a global text processing model that meets the conditions is obtained in a certain round, the model can be sent to Party A, Party B and Party C. After Party A, Party B and Party C deploy the model, if the question text entered by the user is received, the trained text processing model can predict the corresponding answer text and feed it back to the user. For example, the user enters the question text as "Who is the author of the ancient poem ***", and after entering it into the model, Author A can be obtained, and Author A can be fed back to the user.

[0067] In order to avoid the collapse problem of sample distribution, the model optimization method provided in this embodiment can perform a model optimization stage on the client side, first build a business prediction model according to the global model parameters sent by the server, and the business prediction model is used to process multimedia data processing tasks; on this basis, the benchmark data representation and sample distribution information corresponding to the business prediction model can be determined, and local sample data matching the sample distribution information can be selected from the local sample set to form an initial sample set; the sampled initial sample data can be consistent with the sample distribution information, thereby avoiding the collapse of the representation, thereby effectively improving the performance of the client side, and making full use of the complete representation space to obtain rich and diverse distinguishable representation information. On this basis, considering that the initial sample data obtained by sampling according to the sample distribution information has the problem of category separation, that is, samples of the same category are over-dispersed and lose similarity; therefore, the sample distribution difference information can be determined according to the sample data representation and the benchmark data representation corresponding to the initial sample set, and the sample distribution constraint conditions can be calculated according to the sample distribution difference information; at this time, the initial sample set is adjusted by introducing sample distribution constraints, and then the target sample set can be obtained; in the sample sampling stage, a sample set that eliminates the above problems can be obtained, and finally the target sample set can be used to optimize the business prediction model, and the model parameters of the optimized business prediction model are fed back to the server. The server cooperates with multiple clients to complete federated learning, which can avoid the cascading effect caused by the collapse of the representation, thereby realizing multi-party joint training to obtain a business prediction model that meets the usage needs and has stronger generalization ability for use in multimedia data processing tasks to improve the prediction effect.

[0068] Corresponding to the above method embodiment, this specification also provides a model optimization device embodiment, Figure 3 FIG. 1 shows a schematic diagram of a model optimization device provided by an embodiment of the present specification. Figure 3 As shown, the device is applied to a client and includes: The acquisition module 302 is configured to acquire the global model parameters sent by the server, and construct a service prediction model according to the global model parameters, wherein the service prediction model is used to process multimedia data processing tasks; The determination module 304 is configured to determine the benchmark data representation and sample distribution information corresponding to the business prediction model, and select local sample data matching the sample distribution information from the local sample set to form an initial sample set; A calculation module 306 is configured to determine sample distribution difference information according to the sample data representation corresponding to the initial sample set and the benchmark data representation, and calculate sample distribution constraint conditions according to the sample distribution difference information; An adjustment module 308 is configured to adjust the initial sample set according to the sample distribution constraint to obtain a target sample set; The optimization module 310 is configured to optimize the business prediction model using the target sample set, and feed back the model parameters of the optimized business prediction model to the server.

[0069] In an optional embodiment, the acquisition module 302 is further configured to: Obtain global model parameters issued by the server, wherein the global model parameters are obtained by aggregating local model parameters submitted by multiple clients by the server; load an initial business prediction model, and use the global model parameters to update the initial business prediction model to obtain a business prediction model; wherein the multimedia data processing task includes a text processing task, an image processing task, a video processing task or an audio processing task.

[0070] In an optional embodiment, the device further comprises: An information determination module is configured to determine spherical Gaussian distribution information preset for the business prediction model, and use the spherical Gaussian distribution information as sample distribution information; Wherein, the determination module 304 is further configured to: Determine a local sample set corresponding to the business prediction model, and select multiple local sample data from the local sample set according to the spherical Gaussian distribution information; construct an initial sample set based on the multiple local sample data, wherein the distribution information corresponding to the initial sample set is the same as the spherical Gaussian distribution information.

[0071] In an optional embodiment, the calculation module 306 is further configured to: Determine the sample data representation corresponding to the initial sample set; determine the sample distribution difference information by comparing the sample data representation and the benchmark data representation; calculate the sample distribution constraint divergence according to a preset optimization strategy and the sample distribution difference information, and use the sample distribution constraint divergence as a sample distribution constraint condition.

[0072] In an optional embodiment, the optimization module 310 is further configured to: Select target sample data from the target sample set, and use the target sample data to optimize the business prediction model; detect whether the optimized business prediction model meets the local optimization stop condition; if not, select candidate sample data from the target sample set, use the candidate sample data as the target sample data, and execute the step of optimizing the business prediction model using the target sample data; if so, feed back the model parameters of the optimized business prediction model to the server.

[0073] In order to avoid the collapse problem of sample distribution, the model optimization device provided in this embodiment can perform a model optimization stage on the client side, first build a business prediction model according to the global model parameters sent by the server, and the business prediction model is used to process multimedia data processing tasks; on this basis, the benchmark data representation and sample distribution information corresponding to the business prediction model can be determined, and local sample data matching the sample distribution information can be selected from the local sample set to form an initial sample set; the sampled initial sample data can be consistent with the sample distribution information, thereby avoiding the collapse of the representation, thereby effectively improving the performance of the client side, and making full use of the complete representation space to obtain rich and diverse distinguishable representation information. On this basis, considering that the initial sample data obtained by sampling according to the sample distribution information has the problem of category separation, that is, samples of the same category are over-dispersed and lose similarity; therefore, the sample distribution difference information can be determined according to the sample data representation and the benchmark data representation corresponding to the initial sample set, and the sample distribution constraint conditions can be calculated according to the sample distribution difference information; at this time, the initial sample set is adjusted by introducing sample distribution constraints, and then the target sample set can be obtained; in the sample sampling stage, a sample set that eliminates the above problems can be obtained, and finally the target sample set can be used to optimize the business prediction model, and the model parameters of the optimized business prediction model are fed back to the server. The server cooperates with multiple clients to complete federated learning, which can avoid the cascading effect caused by the collapse of the representation, thereby realizing multi-party joint training to obtain a business prediction model that meets the usage needs and has stronger generalization ability for use in multimedia data processing tasks to improve the prediction effect.

[0074] The above is a schematic scheme of a model optimization device of this embodiment. It should be noted that the technical scheme of the model optimization device and the technical scheme of the above-mentioned model optimization method belong to the same concept, and the details not described in detail in the technical scheme of the model optimization device can be referred to the description of the technical scheme of the above-mentioned model optimization method.

[0075] Corresponding to the above method embodiment, this specification also provides a model optimization system embodiment, Figure 4 FIG. 1 shows a schematic diagram of a model optimization system provided by an embodiment of the present specification. Figure 4 As shown, the model optimization system 400 includes a server 410 and multiple clients, including: The target client 420 among the multiple clients is used to receive the global model parameters sent by the server, and build a business prediction model according to the global model parameters, wherein the business prediction model is used to process multimedia data processing tasks; determine the benchmark data representation and sample distribution information corresponding to the business prediction model, and select local sample data matching the sample distribution information in the local sample set to form an initial sample set; determine sample distribution difference information according to the sample data representation corresponding to the initial sample set and the benchmark data representation, and calculate sample distribution constraints according to the sample distribution difference information; adjust the initial sample set according to the sample distribution constraints to obtain a target sample set; optimize the business prediction model using the target sample set, and feed back the model parameters of the optimized business prediction model to the server; The server 410 is used to receive model parameters uploaded by each client, build a target global business prediction model that meets the training stop condition based on the model parameters uploaded by each client, and send the target global model parameters corresponding to the target global business prediction model to each client.

[0076] In an optional embodiment, the acquiring of the global model parameters sent by the server and constructing the business prediction model according to the global model parameters includes: Obtain global model parameters issued by the server, wherein the global model parameters are obtained by aggregating local model parameters submitted by multiple clients by the server; load an initial business prediction model, and use the global model parameters to update the initial business prediction model to obtain a business prediction model; wherein the multimedia data processing task includes a text processing task, an image processing task, a video processing task or an audio processing task.

[0077] In an optional embodiment, before the step of selecting local sample data matching the sample distribution information in the local sample set to form an initial sample set is performed, the method further includes: Determine the spherical Gaussian distribution information preset for the business prediction model, and use the spherical Gaussian distribution information as sample distribution information; wherein, the local sample data matching the sample distribution information is selected in the local sample set to form an initial sample set, including: determining the local sample set corresponding to the business prediction model, and selecting multiple local sample data in the local sample set according to the spherical Gaussian distribution information; constructing an initial sample set based on the multiple local sample data, wherein the distribution information corresponding to the initial sample set is the same as the spherical Gaussian distribution information.

[0078] In an optional embodiment, determining sample distribution difference information according to the sample data representation corresponding to the initial sample set and the benchmark data representation, and calculating sample distribution constraint conditions according to the sample distribution difference information, includes: Determine the sample data representation corresponding to the initial sample set; determine the sample distribution difference information by comparing the sample data representation and the benchmark data representation; calculate the sample distribution constraint divergence according to a preset optimization strategy and the sample distribution difference information, and use the sample distribution constraint divergence as a sample distribution constraint condition.

[0079] In an optional embodiment, the optimizing the business prediction model by using the target sample set and feeding back the model parameters of the optimized business prediction model to the server includes: Select target sample data from the target sample set, and use the target sample data to optimize the business prediction model; detect whether the optimized business prediction model meets the local optimization stop condition; if not, select candidate sample data from the target sample set, use the candidate sample data as the target sample data, and execute the step of optimizing the business prediction model using the target sample data; if so, feed back the model parameters of the optimized business prediction model to the server.

[0080] In order to avoid the collapse problem of sample distribution, the model optimization system provided in this embodiment can perform a model optimization stage on the client side, first build a business prediction model according to the global model parameters sent by the server, and the business prediction model is used to process multimedia data processing tasks; on this basis, the benchmark data representation and sample distribution information corresponding to the business prediction model can be determined, and local sample data matching the sample distribution information can be selected from the local sample set to form an initial sample set; the sampled initial sample data can be consistent with the sample distribution information, thereby avoiding the collapse of the representation, thereby effectively improving the performance of the client side, and making full use of the complete representation space to obtain rich and diverse distinguishable representation information. On this basis, considering that the initial sample data obtained by sampling according to the sample distribution information has the problem of category separation, that is, samples of the same category are over-dispersed and lose similarity; therefore, the sample distribution difference information can be determined according to the sample data representation and the benchmark data representation corresponding to the initial sample set, and the sample distribution constraint conditions can be calculated according to the sample distribution difference information; at this time, the initial sample set is adjusted by introducing sample distribution constraints, and then the target sample set can be obtained; in the sample sampling stage, a sample set that eliminates the above problems can be obtained, and finally the target sample set can be used to optimize the business prediction model, and the model parameters of the optimized business prediction model are fed back to the server. The server cooperates with multiple clients to complete federated learning, which can avoid the cascading effect caused by the collapse of the representation, thereby realizing multi-party joint training to obtain a business prediction model that meets the usage needs and has stronger generalization ability for use in multimedia data processing tasks to improve the prediction effect.

[0081] The above is a schematic scheme of a model optimization system of this embodiment. It should be noted that the technical scheme of the model optimization system and the technical scheme of the above-mentioned model optimization method belong to the same concept, and the details not described in detail in the technical scheme of the model optimization system can be referred to the description of the technical scheme of the above-mentioned model optimization method.

[0082] Figure 5 The structure block diagram of a computing device 500 provided according to an embodiment of the present specification is shown. The components of the computing device 500 include but are not limited to a memory 510 and a processor 520. The processor 520 is connected to the memory 510 via a bus 530, and the database 550 is used to store data.

[0083] The computing device 500 also includes an access device 540 that enables the computing device 500 to communicate via one or more networks 560. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 540 may include one or more of any type of network interface (e.g., a network interface card (NIC)) that is wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a world-wide interoperability for microwave access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, and a near field communication (NFC).

[0084] In one embodiment of the present specification, the above components of the computing device 500 and Figure 5 Other components not shown in the figure may also be connected to each other, for example, via a bus. It should be understood that Figure 5 The computing device structure block diagram shown is only for the purpose of illustration, and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0085] The computing device 500 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smart phone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or a personal computer (PC). The computing device 500 may also be a mobile or stationary server.

[0086] The processor 520 is used to execute the following computer executable instructions, which, when executed by the processor, implement the steps of the above-mentioned model optimization method.

[0087] The above is a schematic scheme of a computing device of this embodiment. It should be noted that the technical scheme of the computing device and the technical scheme of the above-mentioned model optimization method belong to the same concept, and the details not described in detail in the technical scheme of the computing device can be referred to the description of the technical scheme of the above-mentioned model optimization method.

[0088] An embodiment of the present specification also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the above-mentioned model optimization method.

[0089] The above is a schematic scheme of a computer-readable storage medium of this embodiment. It should be noted that the technical scheme of the storage medium and the technical scheme of the above-mentioned model optimization method belong to the same concept, and the details not described in detail in the technical scheme of the storage medium can be referred to the description of the technical scheme of the above-mentioned model optimization method.

[0090] An embodiment of the present specification also provides a computer program product, including a computer program or instructions, which implement the steps of the above-mentioned model optimization method when executed by a processor.

[0091] The above is a schematic scheme of a computer program product of this embodiment. It should be noted that the technical scheme of the computer program product and the technical scheme of the above-mentioned model optimization method belong to the same concept, and the details not described in detail in the technical scheme of the computer program product can be referred to the description of the technical scheme of the above-mentioned model optimization method.

[0092] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0093] The computer instructions include computer program codes, which may be in source code form, object code form, executable files or some intermediate forms, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0094] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of this specification are not limited by the order of the actions described, because according to the embodiments of this specification, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of this specification.

[0095] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0096] The preferred embodiments of this specification disclosed above are only used to help explain this specification. The optional embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of the embodiments of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of this specification, so that technicians in the relevant technical field can understand and use this specification well.

Claims

1. A model optimization method, It is characterized in that Applied to the client, including: Acquire global model parameters sent by the server, and construct a business prediction model according to the global model parameters, wherein the business prediction model is used to process multimedia data processing tasks; Determine the benchmark data representation and sample distribution information corresponding to the business prediction model, and select local sample data matching the sample distribution information from the local sample set to form an initial sample set; Determine sample distribution difference information according to the sample data representation corresponding to the initial sample set and the benchmark data representation, and calculate sample distribution constraint conditions according to the sample distribution difference information; Adjusting the initial sample set according to the sample distribution constraint to obtain a target sample set; The business prediction model is optimized using the target sample set, and the model parameters of the optimized business prediction model are fed back to the server.

2. The model optimization method according to claim 1, It is characterized in that The acquiring of the global model parameters sent by the server and constructing the business prediction model according to the global model parameters includes: Acquire global model parameters sent by the server, wherein the global model parameters are obtained by the server aggregating local model parameters submitted by multiple clients; Loading an initial business prediction model, and updating the initial business prediction model using the global model parameters to obtain a business prediction model; The multimedia data processing tasks include text processing tasks, image processing tasks, video processing tasks or audio processing tasks.

3. The model optimization method according to claim 1, It is characterized in that Before the step of selecting local sample data matching the sample distribution information in the local sample set to form an initial sample set is performed, the method further includes: Determine spherical Gaussian distribution information preset for the business prediction model, and use the spherical Gaussian distribution information as sample distribution information; The step of selecting local sample data matching the sample distribution information from the local sample set to form an initial sample set includes: Determine a local sample set corresponding to the business prediction model, and select a plurality of local sample data from the local sample set according to the spherical Gaussian distribution information; An initial sample set is constructed based on the multiple local sample data, wherein distribution information corresponding to the initial sample set is the same as the spherical Gaussian distribution information.

4. The model optimization method according to claim 1, It is characterized in that The determining of sample distribution difference information according to the sample data representation corresponding to the initial sample set and the benchmark data representation, and calculating the sample distribution constraint condition according to the sample distribution difference information, includes: Determining a sample data representation corresponding to the initial sample set; Determining sample distribution difference information by comparing the sample data representation and the benchmark data representation; The sample distribution constraint divergence is calculated according to a preset optimization strategy and the sample distribution difference information, and the sample distribution constraint divergence is used as a sample distribution constraint condition.

5. The model optimization method according to claim 1, It is characterized in that The optimizing the business prediction model by using the target sample set and feeding back the model parameters of the optimized business prediction model to the server includes: Selecting target sample data from the target sample set, and optimizing the business prediction model using the target sample data; Check whether the optimized business prediction model meets the local optimization stop condition; If not, selecting candidate sample data from the target sample set, using the candidate sample data as the target sample data, and performing a step of optimizing the business prediction model using the target sample data; If so, the model parameters of the optimized business prediction model are fed back to the server.

6. A model optimization device, It is characterized in that Applied to the client, including: An acquisition module is configured to acquire global model parameters sent by a server, and construct a business prediction model according to the global model parameters, wherein the business prediction model is used to process multimedia data processing tasks; A determination module is configured to determine the benchmark data representation and sample distribution information corresponding to the business prediction model, and select local sample data matching the sample distribution information from the local sample set to form an initial sample set; A calculation module, configured to determine sample distribution difference information according to the sample data representation corresponding to the initial sample set and the benchmark data representation, and calculate sample distribution constraint conditions according to the sample distribution difference information; An adjustment module is configured to adjust the initial sample set according to the sample distribution constraint to obtain a target sample set; The optimization module is configured to optimize the business prediction model using the target sample set and feed back the model parameters of the optimized business prediction model to the server.

7. A model optimization system, It is characterized in that The model optimization system includes a server and multiple clients, including: A target client among the multiple clients is used to receive the global model parameters sent by the server, and to construct a business prediction model according to the global model parameters, wherein the business prediction model is used to process multimedia data processing tasks; determine the benchmark data representation and sample distribution information corresponding to the business prediction model, and select local sample data matching the sample distribution information in the local sample set to form an initial sample set; determine sample distribution difference information according to the sample data representation corresponding to the initial sample set and the benchmark data representation, and calculate sample distribution constraints according to the sample distribution difference information; adjust the initial sample set according to the sample distribution constraints to obtain a target sample set; optimize the business prediction model using the target sample set, and feed back the model parameters of the optimized business prediction model to the server; The server is used to receive model parameters uploaded by each client, build a target global business prediction model that meets the training stop condition based on the model parameters uploaded by each client, and send the target global model parameters corresponding to the target global business prediction model to each client.

8. A computing device, It is characterized in that include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method described in any one of claims 1 to 5 are implemented.

9. A computer-readable storage medium, It is characterized in that It stores computer executable instructions, which, when executed by a processor, implement the steps of the method described in any one of claims 1 to 5.

10. A computer program product, It is characterized in that The method comprises a computer program or an instruction, which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 5.

Images