Passive field adaptive method based on active learning

Abstract

According to a passive domain self-adaption method based on active learning, a distribution alignment strategy for fitting a source domain and a target domain in passive domain self-adaption is established, a feature extraction method for clustering nearest neighbor active learning of target domain information is designed, and a target domain learning task is achieved. The method is characterized by comprising the following steps: training a source model by using source domain data and generating a source prototype from source data; then marking target data by using a source prototype, designing a target domain feature extraction method of clustering nearest neighbor active learning, screening out part of data of a target sample, and carrying out label annotation, so that the model can better understand the internal structure of a target domain, fully utilize the information of the target domain, and reduce the cognitive gap of the model between the two domains; and finally, performing fine tuning and updating on the source model by utilizing the marked target sample, realizing inter-domain distribution alignment, and enhancing the classification performance of the model. The invention aims to provide a passive field adaptive method based on active learning, which solves the problems of field distribution offset, category label offset, data privacy security, overhigh cost of data storage and transmission and the like of the existing field adaptive method, enhances the extraction of cross-domain features of data of a source domain and a target domain, and improves the accuracy of the extraction of the cross-domain features of the data of the source domain and the target domain. According to the method, the identification precision of known categories of a source domain is improved, the discovery capability of the model for unknown categories is improved, indirect utilization of source domain data in a pre-training stage is realized, privacy protection and data security protection of the source domain data are promoted, and unsupervised classification learning is performed on unlabeled data of a target domain.

Description

Technical Field

[0001] The present invention belongs to the technical field of domain adaptive pattern classification, and relates to a passive domain adaptation method based on active learning. Background Art

[0002] Machine learning is an important research direction in the era of artificial intelligence. Its research results are widely applied to various fields and play an increasingly important role in people's daily lives. Current machine learning technologies usually rely on closed environments, and their learning processes involve three basic assumptions [1] : One is the independent and identically distributed assumption: the training and test data have the same feature space and data distribution; the second is the closed class label space: the training and test data class spaces are the same; the third is the big data assumption: there are enough available labeled training samples to ensure learning a good classification model. However, the machine learning process in an open environment does not satisfy the above three basic assumptions. Machine learning in a learning environment that does not satisfy the above three assumptions is called machine learning in an open environment. When these three assumption conditions are not met, it is necessary to study the machine learning problem in an open environment, and making machine learning methods adapt to the open environment has become one of the key problems to be solved in the research of the new generation of artificial intelligence technologies.

[0003] Domain Adaptation (DA [2][3] ) can learn a model with good performance on source domain data and adapt it to target domain data with different distributions or even different class label spaces, and solve the learning tasks of target domain data with unlabeled samples or a small number of labeled samples. DA can adapt to the application scenarios of open environments and is one of the main paradigms of machine learning in open environments [4] , and it is the main way to solve the limitation problems of current machine learning in closed environments.

[0004] In real-world scenarios, source domain data is stored on distributed devices. Due to data privacy and security requirements, source domain data cannot be directly accessed and transmitted. Therefore, to solve these problems, some scholars have proposed a more practical unsupervised domain adaptation task: passive domain adaptation [7] . Passive domain adaptation uses the source model parameters trained on source domain data to replace the source domain data for domain adaptation, which not only improves the data security but also reduces the data transmission volume.

[0005] Existing passive domain adaptation [5] is based on closed-set domain adaptation and migrates the pre-trained source model to the unlabeled target domain without accessing any source data to complete the learning tasks of the target domain. Existing passive domain adaptation methods are divided into two categories: white-box SFDA and black-box SFDA [8], the difference between the two lies in whether the parameters of the pre-trained model are available.

[0006] White-box SFDA has two major branches: data generation-based methods [9] and model fine-tuning-based methods [6]

[11] , among which data generation-based methods mostly focus on mining the source domain distribution information contained in the source classifier, and transform the problem into a traditional unsupervised domain adaptation problem by fitting the source domain distribution for solution. However, these methods ignore the indirect utilization of source domain data in the pre-training stage.

[0007] There are mainly three methods for black-box source-free domain adaptation: the method of target domain division, the method of pseudo-label calibration, and the method of distribution adversarial training. ① Method of target domain division: It divides the target domain into easily adaptable and difficult-to-adapt sub-domains, and then realizes cross-domain adaptation through semi-supervised learning. Yang et al.

[12] divide the target domain into easily adaptable and difficult-to-adapt sub-domains, and use semi-supervised learning and data augmentation techniques to effectively reduce the confirmation bias in cross-domain knowledge distillation. ② Method of pseudo-label calibration: It realizes black-box source-free domain adaptation by calibrating the pseudo-labels generated in the source model. Zhang et al.

[13] generate noisy pseudo-labels through a black-box source model, and adopt category sampling and iterative learning strategies to learn a reliable model from unlabeled target domain data. ③ Method of distribution adversarial training: This method requires adding a third-party dataset as a mediator, and then aligns the third-party data and the target domain data through distribution adversarial training. Shi et al.

[14] use a third-party dataset to query the black-box source model to obtain labels, and align the feature distribution of the third-party dataset with the target domain data through distribution adversarial training, and finally fine-tune the target model on the target domain to achieve cross-domain adaptation.

[0008] In the existing research on unsupervised domain adaptation, there are problems such as the low quality of implicit transfer of source domain knowledge, imperfect domain feature learning, and incomplete solution of the indirect distribution alignment between the target domain and the pseudo-source domain. To solve these problems, the patent proposes an unsupervised domain adaptation method based on active learning. First, the source model is trained using source domain data, and source prototypes are generated from the source data. Then, the source prototypes are used to label the target data, and a target domain feature extraction method based on clustering nearest neighbor active learning is designed to screen out some data of the target samples for label annotation, enabling the model to better understand the internal structure of the target domain, fully utilize the target domain information, reduce the cognitive gap between the two domains, and finally use the labeled target samples to fine-tune and update the source model to achieve inter-domain distribution alignment and enhance the classification performance of the model, thus well solving a series of problems such as data security, data storage, and computational burden, and realizing the unsupervised domain adaptation learning task of the source-free domain. [1] Z.Yao, C.Liu, C.Y.Suen. Towards robust pattern recognition: a review. Proceedings of the IEEE, 2020, 108(6): 894-922.

[0009] [2] S.J.Pan, Q.Yang. A survey on transfer learning. IEEE Transactions On Knowledge Data Engineering. 2010, 22(10): 1345-1359.

[0010] [3] L.Zhang, X.Gao. Transfer adaptation learning: a decade survey. IEEE Transactions on Neural Networks and Learning Systems, 2024, 35(1): 23-44.

[0011] [4] Yuan Xiaotong, Zhang Xuyao, Liu Xi, Cheng Zhen, Liu Chenglin. Research progress on machine learning theory for open environments. Pattern Recognition and Artificial Intelligence, 2023, 36(12): 1059-1071.

[0012]

[0013] [5] H.Xia, H.Zhao, Z.Ding. Adaptive adversarial network for source-free domain adaptation.

[0014] IEEE / CVF International Conference on Computer Vision, 2021: 9010-9019.

[0015] [6] J. Lee, G. Lee. Feature alignment by uncertainty and self-training for source-free unsupervised domain adaptation. Neural Networks, 2023: 682-692.

[0016] [7] J. Liang, D. Hu, J. Feng. Do we really need to access the source data?Source hypothesis transfer for unsupervised domain adaptation. Proceedings of the 37th International Conference on Machine Learning, 2020.

[0017] [8] Y. Fang, P. Yap, W. Lin et al. Source-free unsupervised domain adaptation: a survey.

[0018] Neural Networks. 2024: 106230.

[0019] [9] V. K. Kurmi, V. K. Subramanian, V. P. Namboodiri. Domain impression: a source data free domain adaptation method. IEEE / CVF Winter Conference on Applications of Computer Vision, 2021: 615-625.

[0020]

[10] C. Cui, F. Meng, C. Zhang, Z. Liu, L. Zhu, S. Gong, X. Liu. Adversarial source generation for source-free domain adaptation. IEEE Transactions on Circuits and Systems for Video Technology, 2024, 34(6): 1773-1787.

[0021]

[11] Xia. H, Xia. S, Ding. Z.. Discriminative pattern calibration mechanism for source-free domain adaptation. Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition, 2024.

[0022]

[12] Yang J, Peng X, Wang K, Zhu Z, Feng J, Xie L, You Y. Divide-to-adapt: Mitigating confirmation bias for domain adaptation of black-box predictors. Proceedings of the International Conference on Learning Representations, 2023.

[0023]

[13] Zhang. H, Zhang. Y, Jia. K, Zhang. L. Unsupervised domain adaptation of black-box source models. Proceedings of the British Machine Vision Conference, 2021.

[0024]

[14] Shi.Y,Wu.K,Han.Y,Shao.Y,Li.B,Wu.F.Source-free and black-box domain adaptation via distributionally adversarial training.Pattern Recognition 2023,143:109750. Summary of the Invention

[0026] The object of the present invention is to provide a source-free domain adaptation method based on active learning, which solves problems such as domain distribution shift, class label shift, data privacy and security, and high costs of data storage and transmission in existing domain adaptation methods, enhances the extraction of cross-domain features of source domain and target domain data, improves the recognition accuracy of known classes in the source domain, increases the model's ability to discover unknown classes, and at the same time realizes the indirect utilization of source domain data in the pre-training stage, promotes the privacy protection and data security protection of source domain data, and performs unsupervised classification learning on unlabeled data in the target domain.

[0027] The technical solution adopted by the present invention is

[0028] A source-free domain adaptation method based on active learning, which establishes a distribution alignment strategy for fitting the source domain and the target domain in source-free domain adaptation, designs a feature extraction method for clustering nearest neighbor active learning of target domain information, and realizes the target domain learning task. The feature is that first, the source model is trained with source domain data and source prototypes are generated from the source data; then, the source prototypes are used to label the target data, a feature extraction method for clustering nearest neighbor active learning of the target domain is designed, and part of the data of the target samples is screened out for label annotation, so that the model can better understand the internal structure of the target domain, make full use of the target domain information, and reduce the cognitive gap between the two domains; finally, the source model is fine-tuned and updated with the labeled target samples to achieve inter-domain distribution alignment and enhance the classification performance of the model.

[0029] The beneficial effects of the present invention are

[0030] 1. A source model with better classification performance is obtained by pre-training with source domain data. The method will generate accurate feature prototypes to preserve the source domain knowledge, and at the same time lock the feature prototypes during testing. Without using source domain data and only using target domain data, it can reduce the transmission of source domain data and protect the security and privacy of source domain data.

[0031] 2. Constructing a clustering nearest neighbor active learning method can extract effective features of the target domain, screen out part of the data of the target samples for label annotation, and designing active learning can achieve higher accuracy with fewer training labels. By selecting the most valuable samples for labeling and adding them to the training set, the accuracy and efficiency of the model are improved.

[0032] 3. Add the target data with existing labels to the training set to retrain the model, fine-tune the source model, enhance the model's ability to recognize difficult samples, improve the model's generalization ability. At the same time, design cross-entropy loss and target classification loss to optimize the model to further enhance the model's classification performance for the target data. Finally, achieve high-performance domain adaptation learning. Brief Description of the Drawings

[0033] Figure 1 It is the overall framework diagram of the method of the present invention. Detailed Embodiments

[0034] The present invention will be described in detail below in conjunction with the drawings and specific embodiments.

[0035] The general process of the method is divided into three stages: source domain model pre-training, active learning, and domain adaptation.

[0036] (1) First is the source domain model pre-training. Use the source domain data to train the source model and generate source prototypes from the source data. Use the source domain data for pre-training to obtain a source model with better classification performance. Calculate the typical representative source prototype of each category of the source domain data, and then use the source prototype to label the target data.

[0037] (2) Then is to perform active learning. Active learning aims to achieve higher accuracy with fewer training labels. By selecting the most valuable samples for labeling and adding them to the training set, the accuracy and efficiency of the model can be improved. In the SFDA scenario, the target domain cannot access the labeled source domain data, so selective labeling of the target domain data is required. First, cluster the target data to obtain the cluster centers, and determine the pseudo-labels of each cluster center by calculating the similarity between the cluster centers and the source prototypes. To eliminate domain differences and learn the invariant representation of the target domain and improve the model's generalization ability, calculate the nearest neighbors of each cluster center. Weigh the nearest neighbors of each cluster center, and balance the uncertainty and diversity of the selected labels through the uncertainty loss to ensure that the information provided by the selected labels is not redundant and the diversity of the labeled samples. Finally, construct a clustering nearest neighbor active learning method to extract the effective features of the target domain and screen out the target data that meet the requirements.

[0038] (3) Finally is the domain adaptation stage. In this stage, the model will be fine-tuned and updated to achieve domain adaptation. The specific process is as follows: First, add the target data with existing labels to the training set to retrain the model, fine-tune the source model, use the labeled target samples to fine-tune and update the source model, enhance the model's ability to recognize difficult samples, improve the model's generalization ability, then design cross-entropy loss and target classification loss to optimize and evolve the model, further enhance the model's classification performance for the target domain data, and finally design a classifier to complete the domain adaptation learning task.

Claims

1. Generate prototype contrast clustering and three-branch attention passive general domain adaptation method, characterized by The first step is to pre-train the source domain model. The source model is trained using source domain data and the source prototype is generated from the source data. The source model with good classification performance is pre-trained using source domain data. The typical representative source prototype of each category of the source domain data is calculated, and then the target data is labeled using the source prototype.

2. Generate prototype contrast clustering and three-branch attention passive general domain adaptation method, characterized by Then there is active learning; Active learning aims to achieve higher accuracy with fewer training labels, and improves the accuracy and efficiency of the model by selecting the most valuable samples for labeling and adding them to the training set. In the SFDA scenario, the target domain cannot access the labeled source domain data, so the target domain data needs to be selectively labeled. First, the target data is clustered to obtain cluster centers, and the pseudo-labels of each cluster center are determined by calculating the similarity between the cluster centers and the source prototypes. In order to eliminate domain differences and learn invariant representations of the target domain and improve the generalization ability of the model, the nearest neighbors of each cluster center are calculated. The nearest neighbors of each cluster center are weighted, and the uncertainty and diversity of the selected labels are weighed through uncertainty loss to ensure that the information provided by the selected labels is not redundant and the diversity of labeled samples is ensured. Finally, a cluster nearest neighbor active learning method is constructed to extract effective features of the target domain and screen out target data that meets the requirements.

3. Generate prototype contrast clustering and three-branch attention passive general domain adaptation method, characterized by Finally, fine-tune and update the model to achieve domain adaptation; In this stage, we first add the labeled target data to the training set to retrain the model, fine-tune the source model, and use the labeled target samples to fine-tune and update the source model to enhance the model's recognition ability for difficult samples and improve the model's generalization ability. Then, we design the cross entropy loss and target classification loss to optimize the model evolution, further enhance the model's classification performance for target domain data, and finally design a classifier to complete the domain adaptive learning task. Firstly, a feature matching algorithm based on feature distance is designed, and feature prototypes are used to match the target features after class recognition. Then, the cosine similarity between the target feature and the generated prototype feature vector in the same category is calculated, and the most similar feature pairing is found for the target feature of each common class as the input of the domain adaptation part; finally, for samples of the target unknown class, only the target branch is trained.

4. Generate prototype contrast clustering and three-branch attention passive general domain adaptation method, characterized by Finally, a three-branch attention module is constructed to calculate the attention weights of the source domain, target domain, and cross-domain respectively. Enhanced feature alignment effect extracts invariant features between domains, filters pseudo-label noise, realizes the alignment of target domain and pseudo-source domain distribution, and completes cross-domain feature extraction and target domain classification. Specifically, a three-branch attention module is constructed in combination with a classifier to realize target domain feature extraction and cross-domain classification. The three-branch attention module includes two self-attention modules and one cross-attention module. The two self-attention modules extract generated features and target features respectively, while the cross-attention module is used to extract fusion features and assign lower weights to feature pairs with inconsistent pseudo-labels to alleviate the problem of pseudo-label noise. At the same time, distillation loss is introduced to use the output of the cross-attention module branch to guide the training of the target branch, and only the target branch is retained during testing.

Images