AUC-oriented field adaptive method

Abstract

The invention discloses an AUC-oriented domain adaptive method, which comprises the following steps that a neural network model is set, the neural network model is provided with a feature extractor, a main predictor, an auxiliary predictor and a generalization boundary, the feature extractor carries out feature extraction on input data, the main predictor carries out identification on extracted features to obtain a main identification result, and the auxiliary predictor carries out identification on the extracted features to obtain an auxiliary identification result; the auxiliary predictor identifies the extracted features to obtain an auxiliary identification result, and generalizes the boundary to represent the risk of the model; training a neural network model by adopting source domain data, constructing a generalization boundary in the training process, enabling the generalization boundary to represent the risk of the model in combination with the main identification result and the auxiliary identification result, and enabling the risk of the model to be minimized through setting a target function, thereby obtaining the parameters of the neural network model; input data is identified through a neural network model, and a classification result of the data is output by a main predictor. According to the AUC-oriented field adaptive method disclosed by the invention, the recognition accuracy of the model on a new target domain is greatly improved.

Description

Technical Field

[0001] The invention relates to an AUC-guided domain adaptation method, and belongs to the technical field of computer vision. Background Art

[0002] Due to different data sources, environmental conditions, or data collection methods, training data and test data often have differences. This difference can cause the model trained on the source domain (the domain to which the training data belongs) to perform poorly on the target domain (the domain to which the test data belongs).

[0003] In order to solve the problem of mismatch between the distribution of training data and test data, domain adaptation (DA) is proposed in the prior art, which improves the generalization performance of the model in the target domain by utilizing the existing knowledge in the source domain. Its basic assumption is that the source domain and the target domain are similar in some shared features. By transferring the knowledge learned in the source domain to the target domain, the generalization error in the target domain can be effectively reduced. Traditional machine learning methods usually assume that the training and test data are independent and identically distributed, so they cannot handle the domain adaptation problem well. The domain adaptation method reduces the difference between the domains by establishing a mapping relationship between the source domain and the target domain or adjusting the model parameters, thereby improving the performance of the model in the target domain.

[0004] However, although such methods have achieved many results, there is still a gap between theory and practice. Current DA methods usually adopt the paradigm of minimizing the error rate, and their performance on imbalanced data is not satisfactory.

[0005] Therefore, it is necessary to further study domain adaptation methods to solve the above problems. Summary of the invention

[0006] In order to overcome the above problems, the inventors conducted in-depth research and designed an AUC-guided domain adaptation method, which includes the following steps:

[0007] A neural network model is set, wherein the neural network model has a feature extractor, a primary predictor, a secondary predictor, and a generalization bound, wherein:

[0008] The feature extractor extracts features from the input data.

[0009] The main predictor identifies the extracted features and obtains the main identification result.

[0010] The auxiliary predictor identifies the extracted features and obtains auxiliary identification results.

[0011] The generalization bound, which characterizes the risk of the model;

[0012] The neural network model is trained using source domain data, and a generalization bound is constructed during the training process. The generalization bound can characterize the risk of the model by combining the primary identification results and the auxiliary identification results.

[0013] By setting the objective function, the model risk is minimized and the parameters of the neural network model are obtained;

[0014] The input data is identified through the neural network model, and the classification results of the data are output by the auxiliary predictor.

[0015] In a preferred embodiment, the main predictor and the auxiliary predictor have the same structure but different hyperparameters.

[0016] In a preferred embodiment, the construction of the generalization boundary includes the following sub-steps:

[0017] Construct a difference metric to characterize the difference between the source domain dataset and the target domain dataset;

[0018] According to the main recognition results and source domain data, obtain the error rate of the model in the source domain;

[0019] Using the difference metric, the error rate of the model on the target domain is obtained based on the error rate of the model on the source domain.

[0020] In a preferred embodiment, the difference metric is expressed as:

[0021]

[0022] in, Indicates the source domain With the target domain The difference measure between represents the auxiliary predictor, f represents the main predictor, represents the hypothesis class, sup is the supremum function, represents the error between the output of the auxiliary predictor and the main predictor in the target domain, Represents the error between the output of the auxiliary predictor and the output of the main predictor in the source domain.

[0023] In a preferred embodiment, the generalization bound is expressed as:

[0024]

[0025]

[0026] in, represents the error between the true label and the output of the main predictor in the target domain, Represents the error between the true label and the output of the main predictor in the source domain.

[0027] In a preferred embodiment, the objective function is set to

[0028]

[0029]

[0030] f g =f·g

[0031] in, is the output of the auxiliary predictor, f g is the output of the main predictor, and g represents the feature extractor.

[0032] In a preferred embodiment, the training is performed using a gradient descent method, in which the following updates are performed:

[0033]

[0034]

[0035]

[0036] The superscript t represents the number of iterations. represents the loss function value, α 1 , α 2 , α 3 Represents the learning rate parameter.

[0037] In a preferred embodiment, a gradient reversal layer is provided before and after the auxiliary predictor.

[0038] The present invention also provides an electronic device, comprising:

[0039] at least one processor; and

[0040] A memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute any of the methods described above.

[0041] The present invention also provides a computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable the computer to execute any of the above methods.

[0042] The beneficial effects of the present invention include:

[0043] (1) Greatly improves the recognition accuracy of the model in the new target domain;

[0044] (2) It has a wide range of applications, especially has better performance in long-tail data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A flow chart of an AUC-guided domain adaptation method according to a preferred embodiment of the present invention is shown. DETAILED DESCRIPTION

[0046] The present invention will be further described in detail below through the accompanying drawings and embodiments. Through these descriptions, the characteristics and advantages of the present invention will become more clear and distinct.

[0047] The word "exemplary" is used exclusively herein to mean "serving as an example, embodiment, or illustration." Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments. Although various aspects of the embodiments are shown in the drawings, the drawings are not necessarily drawn to scale unless otherwise noted.

[0048] Existing domain adaptation methods usually adopt the paradigm of minimizing the error rate, but their performance on imbalanced data is not satisfactory.

[0049] The prior art also provides an AUC (Area Under the Curve) method, which is a commonly used performance metric for evaluating the quality and prediction accuracy of a binary classification model. AUC is the area under the ROC (Receiver Operating Characteristic) curve, so it is also called the area under the ROC curve. The advantage of AUC is that it is insensitive to data distribution and can comprehensively consider the true positive rate and false positive rate of the model, rather than just focusing on a specific classification threshold. This makes AUC particularly suitable for unbalanced data sets or tasks involving cost sensitivity, because it can provide an overall performance evaluation of the model at various thresholds.

[0050] In the present invention, since AUC has the above characteristics, it is proposed to combine AUC with domain adaptation to solve the problem that the performance of existing domain adaptation methods on imbalanced data is not satisfactory.

[0051] However, existing research on AUC optimization assumes that training and test data come from the same distribution, but in real scenarios, the distribution of training data and test data is often different, and the model trained on the source domain is likely to perform poorly on the target domain. The results of the existing unsupervised domain adaptation (UDA) algorithm cannot be directly applied to AUC optimization because the traditional DA algorithm implicitly requires the loss function to be the sum of independent and identically distributed terms, and AUC does not have this property during the calculation process. Therefore, how to optimize AUC in the presence of such domain differences is a difficult problem of the present invention.

[0052] The present invention provides an AUC-guided domain adaptation method to solve the above problem, comprising the following steps:

[0053] S1. Setting a neural network model, wherein the neural network model has a feature extractor, a main predictor, an auxiliary predictor and a generalization bound, wherein:

[0054] The feature extractor extracts features from the input data.

[0055] The main predictor identifies the extracted features and obtains the main identification result.

[0056] The auxiliary predictor identifies the extracted features and obtains auxiliary identification results.

[0057] The generalization bound, which characterizes the risk of the model;

[0058] S2. Use source domain data to train the neural network model. During the training process, construct a generalization boundary. The generalization boundary is combined with the main identification results and the auxiliary identification results to characterize the risk of the model.

[0059] By setting the objective function, the model risk is minimized and the parameters of the neural network model are obtained;

[0060] S3. The trained neural network model is used to identify the input data, and the auxiliary predictor outputs the classification result of the data.

[0061] In the present invention, a dual-stream structure including a main predictor and an auxiliary predictor is proposed, and the main predictor and the auxiliary predictor share a feature extractor.

[0062] According to the present invention, the main predictor and the auxiliary predictor have the same structure but different hyper parameters.

[0063] The output of the main predictor is used to obtain the prediction result of the model in the source domain, thereby obtaining the risk of the model in the source domain. The risk refers to the error rate of the model classification.

[0064] Furthermore, the risk of the model in the source domain can be expressed as: and in, It represents the output result of the features extracted by the feature extractor after passing through the auxiliary predictor, f g It indicates the output result of the features extracted by the feature extractor after passing through the main predictor. express With f g twice the error in the source domain, Represents the true label and f g 4 times the error in the source domain. Function The expression can be expressed as Among them, f a ,f b represents the variable of the function, and l(·) is an arbitrary convex loss function.

[0065] The output of the auxiliary predictor plays the role of a pseudo-label and can approximate the risk on the target domain, i.e. Approximately in, Represents the error between the output of the auxiliary predictor and the main predictor on the target domain. Since the label on the target domain cannot be obtained in unsupervised learning, it is calculated by To approximate ) represents the approximate risk on the target domain, preferably, it is calculated as

[0066]

[0067] Among them, N c represents the total number of categories, i represents different classes, x, represents the input sample pair, Indicates that the prediction result is the event that the input x belongs to the i-th category, Indicates that the prediction result is input For events that do not belong to the i-th class, l(·) is an arbitrary convex loss function, represents the probability difference of the auxiliary predictor for the input sample on the i-th category, It represents the probability difference of the main predictor for the input sample on the i-th category, Represents the number of samples predicted to be of the i-th category. Represents the number of samples that are not predicted to be of the i-th category.

[0068] Further, Among them, θ i is the probability threshold, I[] means when the condition is met, the corresponding class is returned. Indicates that the predictor's predicted probability of input x belonging to the i-th class is greater than θ i .

[0069] In the present invention, the structures of the feature extractor, the main predictor and the auxiliary predictor are not limited, and those skilled in the art can freely select them according to actual needs.

[0070] According to the present invention, in S2, source domain data is used to train a neural network model, and the source domain data refers to a data set with classification labels.

[0071] According to the present invention, the generalization bound is used to characterize the error rate of the model in the target domain.

[0072] The target domain refers to a new dataset that is different from the source domain and has no classification labels.

[0073] For example, in medical image classification and recognition, medical image data usually has specific characteristics and differences. For example, there may be significant differences between data from different hospitals, different equipment, or different populations. For example, the image data set of one medical institution can be used as the source domain, and the image data of another medical institution can be used as the target domain.

[0074] In a preferred embodiment, the construction of the generalization boundary includes the following sub-steps:

[0075] S21, construct a difference metric to characterize the difference between the source domain dataset and the target domain dataset;

[0076] S22, obtaining the error rate of the model in the source domain according to the main recognition result and the source domain data;

[0077] S23. Using the difference metric, the error rate of the model in the target domain is obtained according to the error rate of the model in the source domain.

[0078] According to the existing theory of domain adaptation, the risk in the target domain With high probability there is an upper bound:

[0079]

[0080] in, represents the empirical risk in the source domain, that is, the training error of the model in the source domain, It is a measure of the difference between the source domain and the target domain, indicating the difference in the distribution of image data in the source domain and the target domain used for training. O(g(N)) is a function of the number of samples N. According to the existing AUC theory, the terms of the AUC loss function do not satisfy independence, and the source domain and the target domain are independent data, so they cannot be directly applied to domain adaptation.

[0081] In S21 of the present invention, a new difference metric is constructed to link the source domain and the target domain, so that it is possible to measure the difference between the source domain and the target domain. Specifically, the difference metric is expressed as:

[0082]

[0083] in, Indicates the source domain With the target domain The difference measure between represents the auxiliary predictor, f represents the main predictor, represents the hypothesis class, sup is the supremum function, represents the error between the output of the auxiliary predictor and the main predictor in the target domain, Represents the error between the output of the auxiliary predictor and the output of the main predictor in the source domain.

[0084] In the above difference metric, the additional coefficient set And 2, the traditional difference metric is changed from the triangle inequality constraint to the convexity constraint, so that the difference metric can cover a wider range of loss function categories and improve the applicability of the model.

[0085] In S22, the model is trained using labeled data from the source domain, and its performance is evaluated on the source domain to obtain the risk on the source domain, thereby reflecting the accuracy of the model on the source domain.

[0086] In a preferred embodiment, the generalization bound obtained in S23 is expressed as:

[0087]

[0088]

[0089] in, represents the error between the true label and the output of the main predictor in the target domain, Represents the error between the true label and the output of the main predictor in the source domain.

[0090] According to a preferred embodiment of the present invention, the objective function is set as

[0091]

[0092]

[0093] f g =f·g

[0094] in, is the output of the auxiliary predictor, f g is the output of the main predictor, and g represents the feature extractor.

[0095] In this objective function, the difference metric, that is, the difference between the data distribution of the source domain training set and the test set distribution, is first obtained by maximizing the inner layer, and then the risk upper bound of the model generalization bound is optimized by minimizing the outer layer.

[0096] In a preferred embodiment, the training is performed using a gradient descent method, in which the following updates are performed:

[0097]

[0098]

[0099]

[0100] The superscript t represents the number of iterations. represents the loss function value, α 1 , α 2 , α 3 Represents the learning rate parameter.

[0101] Since the optimization time cost of min-max is high, in the present invention, a gradient reversal layer is provided before and after the auxiliary predictor, and the optimization process is simplified by introducing the gradient reversal layer.

[0102] The gradient reversal layer (GRL) is a structure proposed in Domain-Adversarial Training of Neural Networks in JMLR2016, which will not be described in detail in the present invention.

[0103] According to the present invention, in S3, the input data of other related data sets can be identified by the trained neural network model, and the classification result of the data is output by the auxiliary predictor.

[0104] For example, in the medical field, the image data set of the first medical institution can be used as the source domain for training, and the trained neural network model can be used to identify the image data of the second medical institution to obtain a more accurate image classification of the second medical institution, which helps to improve the accuracy of doctors' diagnosis of diseases. This has important practical application value for realizing cross-institutional and cross-platform adaptive classification of medical images, which helps to improve patients' diagnosis and treatment results and the utilization efficiency of medical resources.

[0105] Various embodiments of the method described above in the present invention can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include: being implemented in one or more computer programs, which may be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general programmable processor, which may receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0106] The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code can be executed entirely on the machine, partially on the machine, partially on the machine as a stand-alone software package and partially on a remote machine, or entirely on a remote machine or server.

[0107] In the context of the present invention, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0108] To provide interaction with a user, the methods and apparatus described herein may be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and a pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices may also be used to provide interaction with the user; for example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form (including acoustic input, voice input, or tactile input).

[0109] The methods and apparatus described herein may be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communications network). Examples of communications networks include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0110] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship between the client and the server is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services ("Virtual Private Server", or "VPS" for short). The server may also be a server of a distributed system, or a server combined with a blockchain.

[0111] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in the present disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and this document is not limited here.

[0112] Example

[0113] Example 1

[0114] The cross-dataset domain adaptation recognition accuracy test was conducted on five long-tail datasets in real-life scenarios. The five datasets are NSFW, ChestXray, Skin, Imagenet-CIFAR, and DomainNet. The ratio of the number of samples in the head class to the tail class is between 10 and 283.6. Each dataset contains two sub-datasets, and the recognition process includes the following steps:

[0115] A neural network model is set, wherein the neural network model has a feature extractor, a primary predictor, a secondary predictor, and a generalization bound, wherein:

[0116] The feature extractor extracts features from the input data.

[0117] The main predictor identifies the extracted features and obtains the main identification result.

[0118] The auxiliary predictor identifies the extracted features and obtains auxiliary identification results.

[0119] The generalization bound, which characterizes the risk of the model;

[0120] The neural network model is trained using source domain data, and a generalization bound is constructed during the training process. The generalization bound can characterize the risk of the model by combining the primary identification results and the auxiliary identification results.

[0121] By setting the objective function, the model risk is minimized and the parameters of the neural network model are obtained;

[0122] The input data is identified through the neural network model, and the classification result of the data is output by the main predictor.

[0123] The main predictor and the auxiliary predictor have the same structure but different hyper parameters.

[0124] The construction of the generalization bound consists of the following sub-steps:

[0125] Construct a difference metric to characterize the difference between the source domain dataset and the target domain dataset;

[0126] According to the main recognition results and source domain data, obtain the error rate of the model in the source domain;

[0127] Using the difference metric, the error rate of the model on the target domain is obtained based on the error rate of the model on the source domain.

[0128] The difference metric is expressed as:

[0129]

[0130] The generalization bound is expressed as:

[0131]

[0132]

[0133] The objective function is set to

[0134]

[0135]

[0136] f g =f·g

[0137] The training is performed using the gradient descent method, in which the following updates are performed:

[0138]

[0139]

[0140]

[0141] A gradient reversal layer is set before and after the auxiliary predictor.

[0142] Comparative Example

[0143] Comparative Example 1

[0144] The same data set as in Example 1 was used to perform the recognition accuracy test, except that CEOnly, AUConly, AUCMOnly, DANN, CDAN, MINENT, MDD, f-DAL-JS, and f-DAL-Pearson methods were used respectively.

[0145] Among them, CEOnly is a neural network that only uses the CE cross entropy function for identification;

[0146] For the AUCOnly method, please refer to the literature Z. Yang, Q. Xu, S. Bao, X. Cao, and Q. Huang, "Learning with multiclass auc: Theory and algorithms," IEEE Transactions on Pattern Analysis and Machine Intelligence, no. To appear, 2021;

[0147] For the AUCMOnly method, see DJHand and RJTill, "Asimple generalization of the area under the roc curve for multiple class classification problems," Machine Learning, vol.45, no.2, pp.171–186, 2001;

[0148] For the DANN method, please refer to the literature Y.Ganin, E.Ustinova, H.Ajakan, P.Germain, H.Larochelle, F.Laviolette, M.Marchand, and V.Lempitsky, "Domain adversarial training of neural networks," The journal of machine learning research, vol.17, no.1, pp.2096–2030, 2016.;

[0149] For the CDAN method, see M. Long, Z. Cao, J. Wang, and MI Jordan, “Conditional adversarial domain adaptation,” in International Conference on Neural Information Processing Systems, 2018, pp. 1647–1657.

[0150] The MINENT method directly applies CE cross entropy to the migration model for identification;

[0151] For the MDD method, please refer to the literature Y.Zhang, T.Liu, M.Long, and M.Jordan, "Bridging theory and algorithm for domain adaptation," in International Conference on Machine Learning, 2019, pp.7404–7413.;

[0152] The f-DAL-JS method can be found in the literature D.Acuna, G.Zhang, M.T.Law, and S.Fidler, "f-domainadversarial learning: Theory and algorithms," arXiv preprint arXiv:2106.11344, 2021.

[0153] For the f-DAL-Pearson method, see D.Acuna, G.Zhang, M.T.Law, and S.Fidler, “f-domainadversarial learning:Theory and algorithms,” arXiv preprint arXiv:2106.11344, 2021.

[0154] The recognition accuracy results of different methods in Example 1 and Comparative Example 1 are shown in Table 1.

[0155] Table 1

[0156]

[0157]

[0158] Among them, N, D, P, R, H, I, T, C, and R represent different sub-datasets. N→D means taking sub-dataset N as the source domain and sub-dataset D as the target domain.

[0159] It can be seen from Table 1 that the method in Example 1 can achieve a higher accuracy rate than the method in Comparative Example 1, and exceeds the methods in Comparative Example 1 in almost all scenarios.

[0160] In the description of the present invention, it should be noted that the terms "upper", "lower", "inner", "outer", "front", "rear", etc. indicate positions or positional relationships based on the working state of the present invention, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", "third", and "fourth" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0161] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0162] The present invention has been described above in conjunction with preferred embodiments, but these embodiments are only exemplary and serve only as an illustration. On this basis, the present invention may be subjected to a variety of substitutions and improvements, all of which fall within the scope of protection of the present invention.

Claims

1. An AUC-guided domain adaptation method, characterized in that: The following steps are involved: A neural network model is set, wherein the neural network model has a feature extractor, a primary predictor, a secondary predictor, and a generalization bound, wherein: The feature extractor extracts features from the input data. The main predictor identifies the extracted features and obtains the main identification result. The auxiliary predictor identifies the extracted features and obtains auxiliary identification results. The generalization bound, which characterizes the risk of the model; The neural network model is trained using source domain data, and a generalization bound is constructed during the training process. The generalization bound can characterize the risk of the model by combining the primary identification results and the auxiliary identification results. By setting the objective function, the model risk is minimized and the parameters of the neural network model are obtained; The input data is identified through the neural network model, and the classification results of the data are output by the auxiliary predictor.

2. The AUC-guided domain adaptation method according to claim 1, characterized in that: The main predictor and the auxiliary predictor have the same structure but different hyper parameters.

3. The AUC-guided domain adaptation method according to claim 1, characterized in that: The construction of the generalization bound includes the following sub-steps: Construct a difference metric to characterize the difference between the source domain dataset and the target domain dataset; According to the main recognition results and source domain data, obtain the error rate of the model in the source domain; Using the difference metric, the error rate of the model on the target domain is obtained based on the error rate of the model on the source domain.

4. The AUC-guided domain adaptation method according to claim 3, characterized in that: The difference metric is expressed as: in, Indicates the source domain With the target domain The difference measure between represents the auxiliary predictor, f represents the main predictor, represents the hypothesis class, sup is the supremum function, represents the error between the output of the auxiliary predictor and the main predictor in the target domain, Represents the error between the output of the auxiliary predictor and the output of the main predictor in the source domain.

5. The AUC-guided domain adaptation method according to claim 3, characterized in that: The generalization bound is expressed as: in, represents the error between the true label and the output of the main predictor in the target domain, Represents the error between the true label and the output of the main predictor in the source domain.

6. The AUC-guided domain adaptation method according to claim 3, characterized in that: The objective function is set to f g =f·g in, is the output of the auxiliary predictor, f g is the output of the main predictor, and g represents the feature extractor.

7. The AUC-guided domain adaptation method according to claim 1, characterized in that: The training is performed using the gradient descent method, in which the following updates are performed: The superscript t represents the number of iterations. represents the loss function value, and α1, α2, and α3 represent the learning rate parameters.

8. The AUC-guided domain adaptation method according to claim 7, characterized in that: A gradient reversal layer is set before and after the auxiliary predictor.

9. An electronic device, comprising: at least one processor; as well as A memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor so that the at least one processor can execute the method of any one of claims 1-8.

10. A computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1-8.

Images