Domain generalization learning method and device based on multi-element neural average, equipment and medium

Abstract

The application relates to the field of domain generalization learning, and provides a domain generalization learning method and device based on a multi-element neural average, equipment and a medium. The method comprises the following steps: in a training stage, transverse sampling and radial sampling are performed on a neural network, and based on a simplified Jensen-Shannon divergence loss function and an integrated agent multi-element measurement, a target domain generalization risk upper limit is iteratively estimated and optimized to obtain an optimized target domain generalization risk upper limit; in a test stage, the transverse sampling and the radial sampling are closed, and based on the optimized target domain generalization risk upper limit, a globally optimized integrated agent is obtained; target domain samples of a target domain are input into the globally optimized integrated agent to obtain a final prediction value. The application aims to solve the problems of model error setting risk and the incapability of being compatible with multi-source domain and single-source domain in the prior art.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the field of domain generalization learning, and in particular, to a domain generalization learning method based on multi-element neural average, a device, equipment and medium. BACKGROUND

[0002] The purpose of domain generalization learning is to improve the generalization ability of an agent in a target domain by adjusting the learning strategy for source domain data. Since the domain generalization learning task assumes that the target domain data is similar to the source domain data but not identically distributed, it is widely used in data-limited machine learning application scenarios such as medical treatment, intelligent driving and multi-modal learning.

[0003] However, the current domain generalization learning generally assumes that there is an agent with good generalization ability in both the source domain and the target domain in the effective hypothesis space that can be reached in the training stage, while ignoring the memory effect and overfitting trend of current machine learning methods represented by deep learning in the source domain, resulting in the risk of model error setting in existing domain generalization methods. In addition, the method for eliminating the distribution difference or risk difference between the source domains in the prior art cannot adapt to the real needs of single-source domain generalization. SUMMARY

[0004] Embodiments of the present application provide a domain generalization learning method based on multi-element neural average, a device, equipment and medium, aiming to solve the problems of model error setting risk and incompatibility with multi-source domain and single-source domain in the prior art.

[0005] The first aspect of the embodiments of the present application provides a domain generalization learning method based on multi-element neural average, comprising:

[0006] By performing horizontal sampling and radial sampling on the neural network, and based on the reduced Jensen-Shannon divergence loss function and the integration agent diversification metric, the target domain generalization risk upper bound is iteratively estimated and optimized to obtain an optimized target domain generalization risk upper bound;

[0007] By closing the horizontal sampling and the radial sampling, based on the optimized target domain generalization risk upper bound, a globally optimized integrated agent is obtained, and a target domain sample of the target domain is input into the globally optimized integrated agent to obtain a final prediction value.

[0008] Optionally, by performing horizontal sampling and radial sampling on the neural network, and based on the reduced Jensen-Shannon divergence loss function and the integration agent diversification metric, the target domain generalization risk upper bound is iteratively estimated and optimized to obtain an optimized target domain generalization risk upper bound, comprising:

[0009] inputting the source domain sample in the source domain into the neural network;

[0010] the transversal sampling of the neural network obtains an ensemble agent, and the radial sampling of the neural network obtains an ensemble neural average;

[0011] the ensemble agent and the ensemble neural average are input into the reduced Jensen-Shannon divergence loss function and the ensemble agent diversification measure to obtain an estimated target domain generalization risk upper bound;

[0012] the estimated target domain generalization risk upper bound is optimized to obtain the optimized target domain generalization risk upper bound.

[0013] Optionally, the transversal sampling of the neural network obtains an ensemble agent, and specifically:

[0014] the parameters of the neural network are transversally sampled by using a dropout variational family, the dropout variational family is closed, and an ensemble about the dropout variational family is obtained as the ensemble agent.

[0015] Optionally, the radial sampling of the neural network obtains an ensemble neural average, and specifically:

[0016] radial sampling is performed on the neural network optimization trajectory, and a sliding average is adopted to obtain a neural average ensemble.

[0017] Optionally, the reduced Jensen-Shannon divergence loss function is used to optimize the empirical risk of a source domain, and is defined as:

[0018]

[0019] where (x, y) is a sample instance, h is an agent on a deep neural network-induced hypothesis space H, A is a support set of a sampling distribution of the sample instance (x, y), h * is a vector component indicating a real class dimension in an agent output vector, h * ∈ (0, 1), 1 {·} is an indicator function.

[0020] Optionally, the ensemble agent diversification measure is used to show diversification encouraging the transversal sampling, and is defined as:

[0021]

[0022] where ρ is an ensemble agent on a deep neural network-induced hypothesis space H, P is an arbitrary measure, E P is an expectation of the measure P, Var(·) is a variance, and l PJSwherein the reduced Jensen-Shannon divergence loss function is defined as

[0023] Optionally, the target domain generalization risk upper bound is defined as

[0024]

[0025] wherein h is an agent on a hypothesis space H induced by a deep neural network, P is the source domain, and Q is the target domain, the reduced Jensen-Shannon divergence loss function l PJS the generalization risk on the target domain Q, the reduced Jensen-Shannon divergence loss function l PJS the generalization risk on the data domain P, D JS (P||Q) is the Jensen-Shannon divergence between the source domain P and the target domain Q.

[0026] The second aspect of the embodiment of the present application provides a domain generalization learning device based on multi-element neural average, comprising:

[0027] a training module configured to estimate and optimize the target domain generalization risk upper bound iteratively by performing lateral sampling and radial sampling on a neural network and based on a reduced Jensen-Shannon divergence loss function and an integrated agent diversification metric, and obtain an optimized target domain generalization risk upper bound.

[0028] a testing module configured to obtain a globally optimized integrated agent based on the optimized target domain generalization risk upper bound by turning off the lateral sampling and the radial sampling, input target domain samples of a target domain to the globally optimized integrated agent, and obtain a final prediction value.

[0029] The training module further comprises:

[0030] an input sub-module configured to input source domain samples in a source domain to the neural network.

[0031] a sampling sub-module configured to perform the lateral sampling on the neural network to obtain an integrated agent and perform the radial sampling on the neural network to obtain an integrated neural average.

[0032] an estimation sub-module configured to input the integrated agent and the integrated neural average to the reduced Jensen-Shannon divergence loss function and the integrated agent diversification metric to obtain an estimated target domain generalization risk upper bound.

[0033] An optimization submodule is configured to optimize the estimated target domain generalization risk upper bound to obtain an optimized target domain generalization risk upper bound.

[0034] The test module further includes:

[0035] A generation submodule is configured to generate a globally optimized integrated agent based on the optimized target domain generalization risk upper bound by closing the lateral sampling and the radial sampling.

[0036] A test submodule is configured to input a target domain sample of a target domain to the globally optimized integrated agent to obtain a final prediction value.

[0037] The third aspect of the embodiments of the present application provides an electronic device, including a memory, a processor and a computer program stored in the memory, and the processor executes the computer program to implement the steps in the domain generalization learning method based on the multi-element neural average.

[0038] The fourth aspect of the embodiments of the present application provides a computer readable storage medium having a computer program / instruction stored thereon, and the computer program / instruction is executed by a processor to implement the steps in the domain generalization learning method based on the multi-element neural average.

[0039] Advantages:

[0040] The present application provides a domain generalization learning method based on multi-element neural average, device, equipment and medium, by proposing a kind of theoretically guaranteed domain generalization learning method based on multi-element neural average suitable for single source domain data, by the way of fast integrated agent, it is realized that the effective hypothesis space is expanded while not adding learning parameters, with the following advantages:

[0041] (1) simultaneously consider in theory and practice combining domain generalization learning and ensemble learning, and for the first time, a principle integrated agent diversification measurement suitable for domain generalization learning is proposed, which weakens the strong hypothesis of effective hypothesis space existing in domain generalization learning, and reduces the risk of model error setting.

[0042] (2) by diversifying optimization and sampling of agent in lateral and radial directions, the effective hypothesis space is expanded, and the optimization process is stabilized by integration, which improves the generalization performance of agent on target domain data in non-iid scenarios.

[0043] (3) in the method provided by the present application, the target domain risk generalization upper bound is compatible with multi-source domain and single-source domain scenarios, and adapts to the real demand of single-source domain generalization.

[0044] (4) in the method provided by the present application, the neural network radial sampling process does not need to be trained multiple times, has fast integration capability, has smaller computational overhead, and has better large-scale scalability. BRIEF DESCRIPTION OF DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the description of the embodiments of the present application will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0046] Figure 1 is a process flow diagram of a domain generalization learning method based on multi-element neural average proposed by an embodiment of the present application;

[0047] Figure 2 is a training stage flowchart diagram proposed by an embodiment of the present application;

[0048] Figure 3 is a testing stage flowchart diagram proposed by an embodiment of the present application;

[0049] Figure 4 is a device schematic diagram of a domain generalization learning method based on multi-element neural average proposed by an embodiment of the present application. DETAILED DESCRIPTION

[0050] The technical solutions of the embodiments of the present application will be described clearly and completely in the following description of the embodiments of the present application with reference to the drawings. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0051] In the related art, domain generalization learning generally assumes that there is an intelligent agent with good generalization ability on both source domain and target domain in the effective hypothesis space that can be reached in the training stage, and ignores the memory effect and overfitting trend of the current machine learning method represented by deep learning on the source domain, resulting in the risk of model error setting in the existing domain generalization method. In addition, the method for eliminating the distribution difference or risk difference between the source domains in the prior art cannot adapt to the real demand of single-source domain generalization.

[0052] Therefore, the embodiments of the present application propose a domain generalization learning method, device, equipment and medium based on multi-element neural average, which proposes a domain generalization learning method based on multi-element neural average with theoretical guarantee and adaptive to single-source domain data, expands the effective hypothesis space without adding learning parameters by quickly integrating the intelligent agent, and can solve the problems of model error setting risk and incompatibility of multi-source domain and single-source domain in the prior art.

[0053] The domain generalization learning method based on multi-element neural average provided in the embodiments of the present application can be applied when the test target domain data is different from the source domain data. Through the domain generalization learning method based on multi-element neural average provided in the present application, the globally optimized integrated agent of the source domain data sample can be better used for the target domain sample to output the prediction value corresponding to the target domain sample, and the effective migration of data is completed. For example, the method provided in the present application can be applied in the medical field. The data of the target domain patient group is different from the data of the source domain patient group in the training stage (such as age, race, and gender). The method provided in the present application is used to train the data of the target domain patient group as a sample, and the globally optimized integrated agent of the patient group obtained can effectively process the data of the target domain patient group.

[0054] In a first aspect, the embodiments of the present application provide a domain generalization learning method based on multi-element neural average, Figure 1 The flowchart of the domain generalization learning method based on multi-element neural average is shown in FIG. 1. Figure 1 As shown in FIG. 1, the method includes the following two stages:

[0055] S101, training stage.

[0056] In the training stage, the upper bound of the target domain generalization risk is estimated and optimized, and the estimation process and the optimization process are iterated.

[0057] Figure 2 The flowchart of the training stage is shown in FIG. 2. Figure 2 As shown in FIG. 2, the training stage is performed according to the following steps:

[0058] S101-1, inputting the source domain sample in the source domain into the neural network.

[0059] In specific implementation, the source domain sample instance (x, y) in the source domain is input into the neural network, and the neural network belongs to an agent h in the hypothesis space H induced by the deep neural network.

[0060] S101-2, transversely sampling the neural network to obtain an integrated agent, and radially sampling the neural network to obtain an integrated neural average.

[0061] In specific implementation, the parameters θ of the neural network are transversely sampled to obtain a plurality of agents by using the dropout variation family of the neural network, the dropout variation is turned off, the integration about the dropout variation family is obtained as the integrated agent h θ .

[0062] At the same time, the parameters θ of the neural network are radially sampled on the optimization trajectory, and the further, the integrated neural average is obtained by the network structure of the neural network The domain generalization method based on slow ensemble learning in the prior art will train the neural network multiple times in the radial optimization trajectory sampling process, the ensemble speed is slow, and the calculation cost is large. The radial average is used on the optimized trajectory, the algorithm performs sliding average on the parameters of the neural network in the optimization process again, and a diversified ensemble agent is approximately obtained, which improves the ensemble speed and reduces the calculation cost while ensuring the calculation accuracy.

[0063] S101-3, based on the simplified Jensen-Shannon divergence loss function and the ensemble agent diversification measure, an estimated target domain generalization risk upper bound is obtained.

[0064] The ensemble agent h θ and the ensemble neural average The simplified Jensen-Shannon divergence loss function and the ensemble agent diversification measure are input.

[0065] The simplified Jensen-Shannon divergence loss function is used to optimize the source domain empirical risk, and is defined as:

[0066]

[0067] Where (x, y) is a sample instance, h is an agent on a hypothesis space H induced by a deep neural network, A is a support set of a sampling distribution of the sample instance (x, y), h * is a vector component in an agent output vector indicating a real class dimension, h * ∈(0, 1), 1 {·} is an indicator function.

[0068] The simplified Jensen-Shannon divergence loss function can effectively establish a connection between the risk function and a simplified Jensen-Shannon divergence. Similar to the original Jensen-Shannon divergence loss function, the simplified Jensen-Shannon divergence loss function satisfies the triangular inequality of the square root risk; the difference is that the corresponding risk function of the simplified Jensen-Shannon divergence loss function is a convex function when the agent is used as a variable. Therefore, the simplified Jensen-Shannon divergence loss function satisfies the triangular function and convexity at the same time, and it can be proved that the target domain square root generalization risk upper bound of the ensemble agent is limited by the average square root generalization risk of different agents.

[0069] The ensemble agent diversification measure is used to display the lateral sampling diversification of the dropout variational family, and is defined as:

[0070]

[0071] where ρ is an ensemble agent on a hypothesis space H induced by a deep neural network, P is an arbitrary measure, E P is the expectation of the measure P, Var(·) is the variance, l PJS is the reduced Jensen-Shannon divergence loss function.

[0072] For the ensemble agent diversity measure, the following inequality holds:

[0073]

[0074] where, is the agent h with respect to the reduced Jensen-Shannon divergence loss function l PJS the generalization risk on the data domain P.

[0075] Based on the average square root generalization risk of different agents, the ensemble agent diversity measure is introduced, and it is further proved that the target domain square root generalization risk upper bound of the ensemble agent can be further tightened.

[0076] For each instance in the dropout variational family ensemble, the ensemble agent diversity measure is explicitly optimized; for each instance in the ensemble neural average, the ensemble agent diversity measure is implicitly compensated, and based on the generalization empirical risk optimized by the reduced Jensen-Shannon divergence loss function, the estimated target domain generalization risk upper bound is obtained.

[0077] The estimated target domain generalization risk upper bound is compatible with multi-source domain and single-source domain scenarios, and adapts to the real needs of single-source domain generalization. The target domain generalization risk upper bound is defined as:

[0078]

[0079] where h is an agent on a hypothesis space H induced by a deep neural network, P is the source domain, Q is the target domain, is the agent ρ with respect to the reduced Jensen-Shannon divergence loss function l PJS the generalization risk on the target domain Q, is the agent h with respect to the reduced Jensen-Shannon divergence loss function l PJS the generalization risk on the data domain P, D JS (P||Q) is the Jensen-Shannon divergence between the source domain P and the target domain Q. For fixed source domain P and target domain Q, the divergence is a fixed value.

[0080] S101-4, optimizing the estimated target domain generalization risk upper bound to obtain an optimized target domain generalization risk upper bound.

[0081] The estimated target domain generalization risk upper bound obtained above is optimized using a gradient descent method, and when convergence is obtained, an optimized target domain generalization risk upper bound is obtained.

[0082] S102, test phase.

[0083] In the test phase, based on the target domain generalization risk upper bound obtained in the training phase, a globally optimized integrated agent is obtained The globally optimized integrated agent is used to process the target domain sample x t to obtain the final prediction value

[0084] Figure 3 The test phase flowchart is shown, as shown in Figure 3 , the test phase is carried out according to the following steps:

[0085] S102-1, turn off the horizontal sampling and the radial sampling, and based on the optimized target domain generalization risk upper bound, obtain a globally optimized integrated agent.

[0086] In specific implementation, after the optimization of the target domain risk upper bound is completed, by turning off the dropout and the neural average, based on the optimized target domain generalization risk upper bound, a globally optimized integrated agent is obtained The globally optimized integrated agent has a controlled target domain generalization risk upper bound.

[0087] S102-2, input the target domain sample of the target domain to the globally optimized integrated agent to obtain the final prediction value.

[0088] The target domain sample x t is input to the globally optimized integrated agent with a controlled target domain generalization risk upper bound to obtain the final prediction value

[0089] For example, the domain generalization learning method based on multi-element neural average provided by the embodiments of the present application is applied to the automatic driving field, the weather distribution data in the source domain is different from the weather distribution data in the target domain, and the weather distribution data in the source domain is trained based on the above-mentioned domain generalization learning method based on multi-element neural average. After the global optimization of the weather distribution integrated agent, the weather distribution data in the target domain can be effectively processed, and the prediction value corresponding to the required target domain weather distribution data is output.

[0090] The embodiment of the application provides a domain generalization learning method based on multi-element neural average.

[0091] (1) The domain generalization learning and ensemble learning are considered in theory and practice, and a principle ensemble agent diversification measurement suitable for domain generalization learning is first proposed, the strong assumption of the effective hypothesis space of the domain generalization learning is weakened, and the model error setting risk is reduced.

[0092] (2) The effective hypothesis space is expanded by optimizing and sampling the agent in the horizontal direction and the radial direction, and the optimization process is stabilized by integration, and the generalization performance of the agent on the target domain data in the non-identically distributed scenario is improved.

[0093] (3) In the method provided by the application, the target domain risk generalization upper bound is compatible with the multi-source domain and the single-source domain situation, and adapts to the real demand of single-source domain generalization.

[0094] (4) In the method provided by the application, the neural network radial sampling process does not need to be trained multiple times, has fast integration capability, smaller calculation overhead, and better large-scale scalability.

[0095] Based on the same inventive concept, the embodiment of the application provides a domain generalization learning device based on multi-element neural average, Figure 4 The domain generalization learning device based on multi-element neural average is shown in the schematic diagram, Figure 4 As shown in the schematic diagram, the domain generalization learning device based on multi-element neural average comprises:

[0096] A training module is configured to estimate and optimize the target domain generalization risk upper bound by performing horizontal sampling and radial sampling on a neural network and based on a simplified Jensen-Shannon divergence loss function and an ensemble agent diversification measurement, and obtain an optimized target domain generalization risk upper bound.

[0097] A test module is configured to obtain a globally optimized ensemble agent based on the optimized target domain generalization risk upper bound by closing the horizontal sampling and the radial sampling, input target domain samples of a target domain to the globally optimized ensemble agent, and obtain a final prediction value.

[0098] The training module further comprises:

[0099] An input sub-module is configured to input source domain samples in a source domain to the neural network.

[0100] The sampling sub-module is configured to perform the lateral sampling on the neural network to obtain an integrated agent, and perform the radial sampling on the neural network to obtain an integrated neural average.

[0101] The estimation sub-module is configured to input the integrated agent and the integrated neural average into the reduced Jensen-Shannon divergence loss function and the integrated agent diversification metric to obtain an estimated target domain generalization risk upper bound.

[0102] The optimization sub-module is configured to optimize the estimated target domain generalization risk upper bound to obtain an optimized target domain generalization risk upper bound.

[0103] The test module further includes:

[0104] The generation sub-module is configured to generate a globally optimized integrated agent based on the optimized target domain generalization risk upper bound by closing the lateral sampling and the radial sampling.

[0105] The test sub-module is configured to input a target domain sample of a target domain into the globally optimized integrated agent to obtain a final prediction value.

[0106] The domain generalization learning device based on the multi-element neural average provided in the embodiments of the present application can implement all the contents of the foregoing method embodiments, and therefore, the specific function implementation of each module can refer to the foregoing method embodiments, which will not be described herein again.

[0107] The embodiments of the present application further provide an electronic device including a memory, a processor, and a computer program stored in the memory, and the processor executes the computer program to implement the steps in the domain generalization learning method based on the multi-element neural average provided in the embodiments of the present application.

[0108] In yet another embodiment provided in the present application, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program / instruction, and the computer program / instruction is executed by a processor to implement the steps in the domain generalization learning method based on the multi-element neural average provided in the embodiments of the present application.

[0109] In the embodiments described above, all or some of the steps can be implemented by hardware, software, firmware or any combination thereof. When implemented by software, all or some of the steps can be implemented in the form of one or more computer programs or program elements. The computer programs reside (at least temporarily) in a memory of a computer during execution. The memory can be a RAM memory, a flash memory, a ROM memory, an EPROM memory, or any other suitable memory. The memory can be integral to or separate from the computer. The computer programs can be written in any suitable programming language, such as C, C++, Java, Visual Basic, etc. The computer programs can be written in assembly or machine language, if desired. The computer programs can be distributed over network coupled file servers, or can be distributed by any other suitable means.

[0110] It is to be noted that, in the present document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily implying any actual relationship or order between such entities or actions. Also, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. In addition, terms such as "first" and "second" are used herein only to distinguish one element from another, without necessarily implying any actual relationship or order between such elements.

[0111] Each of the embodiments described in the present specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other. Each of the embodiments mainly describes the difference from other embodiments. In particular, the system embodiments are described in a relatively simple manner because they are substantially similar to the method embodiments, and the relevant parts can be referred to the description of the method embodiments.

[0112] The above merely provides the preferred embodiments of the application, and not intended to limit the protection scope of the application. Any modification, equivalent replacement, improvement, and the like made within the principle and technical scope of the application shall fall into the protection scope of the application.

Claims

1. A domain generalization learning method based on multivariate neural averaging, characterized in that, include: By performing lateral and radial sampling on the neural network, and based on the simplified Jensen-Shannon divergence loss function and the ensemble agent diversification measure, the upper bound of the target domain generalization risk is estimated and optimized iteratively to obtain the optimized upper bound of the target domain generalization risk. By disabling the lateral sampling and the radial sampling, a global optimization ensemble agent is obtained based on the optimized target domain generalization risk upper bound. The target domain sample of the target domain is then input into the global optimization ensemble agent to obtain the final prediction value. In this context, the domain generalization learning method based on multivariate neural averaging is applied to the field of autonomous driving. The weather distribution data in the source domain is different from the weather distribution data in the target domain. The weather distribution data in the source domain is used as a sample and trained based on the domain generalization learning method based on multivariate neural averaging to obtain a globally optimized weather distribution integrated agent. The globally optimized weather distribution integrated agent is used to process the weather distribution data in the target domain and output the predicted value corresponding to the weather distribution data in the target domain. The simplified Jensen-Shannon divergence loss function, used to optimize the empirical risk of the source domain, is defined as follows: in, Let h be a sample instance, h be an agent on the hypothesis space H induced by the deep neural network, and A be the sample instance. The support set of the sampling distribution The vector component in the agent's output vector that indicates the dimension of the true class. , This is an indicator function.

2. The domain generalization learning method based on multivariate neural averaging as described in claim 1, characterized in that, By performing lateral and radial sampling on the neural network, and based on a simplified Jensen-Shannon divergence loss function and an ensemble agent diversification metric, the upper bound of the target domain generalization risk is iteratively estimated and optimized to obtain the optimized upper bound of the target domain generalization risk, including: Input source domain samples from the source domain into the neural network; The lateral sampling of the neural network yields an integrated agent, and the radial sampling of the neural network yields an integrated neural average. The average inputs of the integrated agent and the integrated neuron to the simplified Jensen-Shannon divergence loss function and the integrated agent diversification measure are used to obtain the estimated upper bound of the target domain generalization risk. The estimated target domain generalization risk upper bound is optimized to obtain the optimized target domain generalization risk upper bound.

3. The domain generalization learning method based on multivariate neural averaging as described in claim 2, characterized in that, The lateral sampling of the neural network yields the integrated agent, specifically as follows: The parameters of the neural network are laterally sampled using a dropout variational family, and the dropout variation is turned off to obtain an ensemble of the dropout variational family, which serves as the ensemble agent.

4. The domain generalization learning method based on multivariate neural averaging as described in claim 2, characterized in that, The radial sampling of the neural network is used to obtain the integrated neural average, specifically as follows: Radial sampling is performed on the optimized trajectory of the neural network, and a moving average is used to obtain the average neural ensemble.

5. The domain generalization learning method based on multivariate neural averaging as described in claim 2, characterized in that, The integrated agent diversity metric, used to display the diversity that encourages the lateral sampling, is defined as: in, Let P be an ensemble agent on a hypothesis space H induced by a deep neural network, where P is an arbitrary measure. The expectation of the measure P, To obtain the variance, The simplified Jensen-Shannon divergence loss function is given.

6. The domain generalization learning method based on multivariate neural averaging as described in claim 5, characterized in that, The upper bound of the target domain generalization risk is defined as follows: Where h is an agent in the hypothesis space H induced by the deep neural network, P is the source domain, and Q is the target domain. For the intelligent agent Regarding the generalization risk of the simplified Jensen-Shannon divergence loss function on the target domain Q... For agent h, the loss function with respect to the simplified Jensen-Shannon divergence is... Generalization risk over data domain P Let be the Jensen-Shannon divergence between the source domain P and the target domain Q.

7. A domain generalization learning device based on multivariate neural averaging, characterized in that, The device includes: The training module is used to estimate and optimize the upper bound of the target domain generalization risk iteratively by performing lateral and radial sampling on the neural network and based on the simplified Jensen-Shannon divergence loss function and the ensemble agent diversification measure, so as to obtain the optimized upper bound of the target domain generalization risk. The testing module is used to obtain a global optimization ensemble agent based on the optimized target domain generalization risk upper bound by disabling the lateral sampling and the radial sampling, input target domain samples of the target domain into the global optimization ensemble agent, and obtain the final prediction value. In this context, the domain generalization learning method based on multivariate neural averaging is applied to the field of autonomous driving. The weather distribution data in the source domain is different from the weather distribution data in the target domain. The weather distribution data in the source domain is used as a sample and trained based on the domain generalization learning method based on multivariate neural averaging to obtain a globally optimized weather distribution integrated agent. The globally optimized weather distribution integrated agent is used to process the weather distribution data in the target domain and output the predicted value corresponding to the weather distribution data in the target domain. The simplified Jensen-Shannon divergence loss function, used to optimize the empirical risk of the source domain, is defined as follows: in, Let h be a sample instance, h be an agent on the hypothesis space H induced by the deep neural network, and A be the sample instance. The support set of the sampling distribution The vector component in the agent's output vector that indicates the dimension of the true class. , This is an indicator function.

8. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps in the domain generalization learning method based on multivariate neural averaging as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps in the domain generalization learning method based on multivariate neural averaging as described in any of claims 1-6.

Images