Unlabeled Target Research Articles

SEA++: Multi-Graph-Based Higher-Order Sensor Alignment for Multivariate Time-Series Unsupervised Domain Adaptation.

Unsupervised Domain Adaptation (UDA) methods have been successful in reducing label dependency by minimizing the domain discrepancy between labeled source domains and unlabeled target domains. However, these methods face challenges when dealing with Multivariate Time-Series (MTS) data. MTS data typically originates from multiple sensors, each with its unique distribution. This property poses difficulties in adapting existing UDA techniques, which mainly focus on aligning global features while overlooking the distribution discrepancies at the sensor level, thus limiting their effectiveness for MTS data. To address this issue, a practical domain adaptation scenario is formulated as Multivariate Time-Series Unsupervised Domain Adaptation (MTS-UDA). In this paper, we propose SEnsor Alignment (SEA) for MTS-UDA, aiming to address domain discrepancy at both local and global sensor levels. At the local sensor level, we design endo-feature alignment, which aligns sensor features and their correlations across domains. To reduce domain discrepancy at the global sensor level, we design exo-feature alignment that enforces restrictions on global sensor features. We further extend SEA to SEA++ by enhancing the endo-feature alignment. Particularly, we incorporate multi-graph-based higher-order alignment for both sensor features and their correlations. Extensive empirical results have demonstrated the state-of-the-art performance of our SEA and SEA++ on six public MTS datasets for MTS-UDA.

Global-Local Continual Transfer Network for Intelligent Fault Diagnosis of Rotating Machinery

Existing fault diagnosis methods face three fundamental challenges when deployed in the dynamic environment, including insufficient continuous diagnostic capability, poor model generalization performance, and inadequate data privacy protection. To address these problems, we develop a novel continual fault diagnosis framework named Global-Local Continual Transfer Network (GLCTN) for fault classification of unlabeled target samples under different working conditions without any source samples. Specifically, a consistency loss and a mutual information loss are introduced in the proposed GLCTN to transfer the learned diagnostic knowledge. Moreover, a dual-speed optimization strategy is utilized to preserve the acquired diagnostic knowledge and to endow the model with the ability to acquire new knowledge. Validation experiments conducted on an automobile transmission dataset demonstrate that the proposed GLCTN achieves satisfactory diagnostic performance on multiple continuous transfer diagnostic tasks.

SAFENet: Semantic-Aware Feature Enhancement Network for unsupervised cross-domain road scene segmentation

Unsupervised cross-domain road scene segmentation has attracted substantial interest because of its capability to perform segmentation on new and unlabeled domains, thereby reducing the dependence on expensive manual annotations. This is achieved by leveraging networks trained on labeled source domains to classify images on unlabeled target domains. Conventional techniques usually use adversarial networks to align inputs from the source and the target in either of their domains. However, these approaches often fall short in effectively integrating information from both domains due to Alignment in each space usually leads to bias problems during feature learning. To overcome these limitations and enhance cross-domain interaction while mitigating overfitting to the source domain, we introduce a novel framework called Semantic-Aware Feature Enhancement Network (SAFENet) for Unsupervised Cross-domain Road Scene Segmentation. SAFENet incorporates the Semantic-Aware Enhancement (SAE) module to amplify the importance of class information in segmentation tasks and uses the semantic space as a new domain to guide the alignment of the source and target domains. Additionally, we integrate Adaptive Instance Normalization with Momentum (AdaIN-M) techniques, which convert the source domain image style to the target domain image style, thereby reducing the adverse effects of source domain overfitting on target domain segmentation performance. Moreover, SAFENet employs a Knowledge Transfer (KT) module to optimize network architecture, enhancing computational efficiency during testing while maintaining the robust inference capabilities developed during training. To further improve the segmentation performance, we further employ Curriculum Learning, a self-training mechanism that uses pseudo-labels derived from the target domain to iteratively refine the network. Comprehensive experiments on three well-known datasets, “Synthia→Cityscapes” and “GTA5→Cityscapes”, demonstrate the superior performance of our method. In-depth examinations and ablation studies verify the efficacy of each module within the proposed method.

Deep joint subdomain alignment for unsupervised domain adaptation

Domain adaptation (DA) aims to transfer knowledge of the labeled source domain to the unlabeled target domain. Current DA methods learn domain-invariant features by aligning source and target feature spaces by distribution discrepancy minimization or adversarial training. However, existing DA methods fail to consider the semantic-level alignment, which can lead to the distortion of semantic feature structures for the target tasks. In this paper, we propose a novel domain adaptation method, called Deep Joint Subdomain Alignment (DJSA), which jointly conducts intra-subdomain and inter-subdomain alignment for fine-grained classification of target tasks. DJSA obtains the pseudo labels of unlabeled target samples by measuring the similarity between the target samples and source category centers, and then DJSA calculates an intra-subdomain loss to reduce the difference within the same class and increase the difference between different classes in both the source domain and the target domain, and DJSA utilizes a Local Maximum Mean Discrepancy (LMMD) metric to measure the inter-subdomain discrepancy between the source and target domain, and minimizes the discrepancy to align different domains. Furthermore, extensive experimental results demonstrate the effectiveness and superiority of our proposed method in unsupervised domain adaptation.

Attention-disentangled re-ID network for unsupervised domain adaptive person re-identification

Unsupervised domain adaptation (UDA) for person re-identification (re-ID) aims to bridge the domain gap by transferring knowledge from the labeled source domain to the unlabeled target domain. Recently, pseudo-label-based approaches have become the dominant solution for addressing this issue. However,most pseudo-label-based methods neglect class boundary samples to reduce the influence of false pseudo labels, sacrificing intra-class semantic diversity. Although hard sample memory bank-based approaches can discover relationships between samples and describe intra-class diversity, they are prone to increasing the risks of generating incorrect pseudo labels. Balancing intra-class variety and pseudo-label accuracy in cross-domain person re-ID poses a challenge. In this study, we propose an attention-disentangled re-ID network (ADDNet) to enhance the discriminative ability of re-ID related feature representations, addressing the contradiction between intra-class diversity and pseudo-label accuracy. Unlike most disentanglement learning-based re-ID approaches that focus on separating explanatory factors, we design a spatial attention-disentangled mechanism to separate re-ID related and unrelated weights, enhancing the discriminative ability of cross-domain feature representation. Additionally, a multiple hard sample memory learning strategy is designed to express intra-class diversity of target samples using re-ID related features extracted from ADDNet. Extensive experiments show that ADDNet achieves 46.7% mAP on the Market-to-MSMT cross-domain re-ID task, outperforming state-of-the-art methods by 6.5 points.

Consistency-guided Multi-Source-Free Domain Adaptation

Deep neural networks suffer from severe performance degradation when facing a distribution shift between the labeled source domain and unlabeled target domain. Domain adaptation addresses this issue by aligning the feature distributions of both domains. Conventional methods assume that the labeled source samples are drawn from a single data distribution (domain) and can be fully accessed during training. However, in real applications, multiple source domains with different distributions often exist, and source samples may be unavailable due to privacy and storage constraints. To address multi-source and data-free challenges, Multi-Source-Free Domain Adaptation (MSFDA) uses only diverse pre-trained source models without requiring any source data. Most existing MSFDA methods adapt each source model to the target domain individually, making them ineffective in leveraging the complementary transferable knowledge from different source models. In this paper, we propose a novel COnsistency-guided multi-source-free Domain Adaptation (CODA) method, which leverages the label consistency criterion as a bridge to facilitate the cooperation among source models. CODA applies consistency regularization on the soft labels of weakly- and strongly-augmented target samples from each pair of source models, allowing them to supervise each other. To achieve high-quality pseudo-labels, CODA also performs a consistency-based denoising to unify the pseudo-labels from different source models. Finally, CODA optimally combines different source models by maximizing the mutual information of the predictions of the resulting target model. Extensive experiments on four benchmark datasets demonstrate the effectiveness of CODA compared to the state-of-the-art methods.

A Novel Fault Diagnosis Method for Rolling Bearings Based on Multi-Domain Adaptation Neural Network

In recent years, data-driven approaches have shown promise in fault diagnosis. However, in practical industrial applications, challenges such as discrepancies between source domain and target domain data distributions, coupled with limited labeled fault data, have hindered the performance of traditional domain adaptation algorithms in bearing fault diagnosis. To address this issue, this study introduces a novel method based on MDANN (Multi Domain Adaptation Neural Network) for diagnosing rolling bearing faults using unlabeled data. Initially, the raw vibration signals are preprocessed using wavelet packet decomposition and reconstruction (WPT), which helps minimize signal redundancy while preserving critical features. Following this, the multi-kernel maximum mean discrepancy (MK-MMD) algorithm is employed to compute the differences in input feature values, and the MDANN network parameters are adjusted via backpropagation, allowing the network to extract domain-invariant features. To ensure the unlabeled target domain data can be effectively used in training, a pseudo-labeling strategy is applied, where the label with the highest probability is treated as the true label, thereby enabling the model to learn from the target domain data and improve the acquisition of reliable diagnostic knowledge. The method is validated using two publicly available datasets, CWRU and PU, with experimental results demonstrating that it outperforms conventional domain adaptation techniques in terms of diagnostic accuracy. This indicates the proposed approach is highly effective in capturing transferable features and addressing the distributional differences between datasets.

FGTL: Federated Graph Transfer Learning for Node Classification

Unsupervised multi-source domain transfer in federated scenario has become an emerging research direction, which can help unlabeled target domain to obtain the adapted model through source domains under privacy-preserving. However, when local data is graph, the difference of domains (or data heterogeneity) mainly originate from the difference in node attributes and subgraph structures, leading to serious model drift, which is not considered by the existing related algorithms. Currently, there are two challenges in this scenario: (1) The node representations extracted directly through conventional GNNs lack inter-domain generalized and consistent information, making it difficult to apply existing federated learning algorithms. (2) The knowledge of source domains has quality differences, which may lead to negative transfer. To address these issues, we propose a novel two-phase Federated Graph Transfer Learning (FGTL) framework. In the generalization phase, FGTL utilizes local contrastive learning and global context embedding to force node representations to capture the inter-domain generalized and consistent information, lightly alleviating model drift. In the transfer phase, FGTL utilizes consensus knowledge to force the decision bound of classifier to adapt to the target client. In addition, FGTL+ exploits model grouping to make consensus knowledge generation more efficient, further enhancing the scalability of FGTL. Extensive experiments show that FGTL significantly outperforms state-of-the-art related methods, while FGTL+ further enhances privacy protection and reduces both communication and computation overhead.

Fault Impulse Inference and Cyclostationary Approximation: A feature-interpretable intelligent fault detection method for few-shot unsupervised domain adaptation

With the advancement of intelligent detection for rotating machinery, numerous domain adaptation methods have been devised to transfer detection knowledge from one source domain working condition to another target domain working condition, involving extensive transfer scenarios including labeled, few-shot labeled, and unlabeled target conditions. Yet, learning from sparsely labeled signals in the source domain working condition and transferring to unlabeled target conditions, termed few-shot unsupervised domain adaptation (FUDA), is closer to reality but almost unexplored. Diverging from the intuition of combining existing transfer and few-shot learning technologies, this paper pioneers a novel single learning principle focusing on the cyclostationary mechanism (CT) of fault signals. In its implementation, named cyclically enhanced cyclostationary variational autoencoder (CCTVAE), the CT principle motivates the encoder to infer domain-shared representations with fault impulses, and the decoder approximates the cyclostationary structure containing the clear fault and working condition information. Then, auxiliary samples for few-shot expansion are generated by adjusting cyclic parameters of the posterior distribution of representations. Experimentally, CCTVAE achieves commendable results on simulated and real fault datasets. Even for compound faults, domain-shared representations and generated auxiliary signals manifest interpretable fault-indicating spectral lines in the frequency domain, underscoring method reliability.

A Collaborative Domain Adversarial Network for Unlabeled Bearing Fault Diagnosis

At present, data-driven fault diagnosis has made significant achievements. However, in actual industrial environments, labeled fault data are difficult to obtain, making the industrial application of intelligent fault diagnosis models very challenging. This limitation even prevents intelligent fault diagnosis algorithms from being applicable in real-world industrial settings. In light of this, this paper proposes a Collaborative Domain Adversarial Network (CDAN) method for the fault diagnosis of rolling bearings using unlabeled data. First, two types of feature extractors are employed to extract features from both the source and target domain samples, reducing signal redundancy and avoiding the loss of critical signal features. Second, the multi-kernel clustering algorithm is used to compute the differences in input feature values, create pseudo-labels for the target domain samples, and update the CDAN network parameters through backpropagation, enabling the network to extract domain-invariant features. Finally, to ensure that unlabeled target domain data can participate in network training, a pseudo-label strategy using the maximum probability label as the true label is employed, addressing the issue of unlabeled target domain data not being trainable and enhancing the model’s ability to acquire reliable diagnostic knowledge. This paper validates the CDAN using two publicly available datasets, CWRU and PU. Compared with four other advanced methods, the CDAN method improved the average recognition accuracy by 7.85% and 5.22%, respectively. This indirectly proves the effectiveness and superiority of the CDAN in identifying unlabeled bearing faults.

An Evidential Multi-Target Domain Adaptation Method Based on Weighted Fusion for Cross-Domain Pattern Classification.

For cross-domain pattern classification, the supervised information (i.e., labeled patterns) in the source domain is often employed to help classify the unlabeled target domain patterns. In practice, multiple target domains are usually available. The unlabeled patterns (in different target domains) which have high-confidence predictions, can also provide some pseudo-supervised information for the downstream classification task. The performance in each target domain would be further improved if the pseudo-supervised information in different target domains can be effectively used. To this end, we propose an evidential multi-target domain adaptation (EMDA) method to take full advantage of the useful information in the single-source and multiple target domains. In EMDA, we first align distributions of the source and target domains by reducing maximum mean discrepancy (MMD) and covariance difference across domains. After that, we use the classifier learned by the labeled source domain data to classify query patterns in the target domains. The query patterns with high-confidence predictions are then selected to train a new classifier for yielding an extra piece of soft classification results of query patterns. The two pieces of soft classification results are then combined by evidence theory. In practice, their reliabilities/weights are usually diverse, and an equal treatment of them often yields the unreliable combination result. Thus, we propose to use the distribution discrepancy across domains to estimate their weighting factors, and discount them before fusing. The evidential combination of the two pieces of discounted soft classification results is employed to make the final class decision. The effectiveness of EMDA was verified by comparing with many advanced domain adaptation methods on several cross-domain pattern classification benchmark datasets.

Pseudo-label refinement via hierarchical contrastive learning for source-free unsupervised domain adaptation

Source-free unsupervised domain adaptation aims to adapt a source model to an unlabeled target domain without accessing the source data due to privacy considerations. Existing works mainly solve the problem by self-training methods and representation learning. However, these works typically learn the representation on a single semantic level and barely exploit the rich hierarchical semantic information to obtain clear decision boundaries, which makes it hard for these methods to achieve satisfactory generalization performance. In this paper, we propose a novel hierarchical contrastive domain adaptation algorithm that exploits self-supervised contrastive learning on both fine-grained instances and coarse-grained cluster semantics. On the one hand, we propose an adaptive prototype pseudo-labeling strategy to obtain much more reliable labels. On the other hand, we propose hierarchical contrastive representation learning on both fine-grained instance-wise level and coarse-grained cluster level to reduce the negative effect of label noise and stabilize the whole training procedure. Extensive experiments are conducted on primary unsupervised domain adaptation benchmark datasets, and the results demonstrate the effectiveness of the proposed method.

Adaptive wavelet-VNet for single-sample test time adaptation in medical image segmentation.

In medical image segmentation, a domain gap often exists between training and testing datasets due to different scanners or imaging protocols, which leads to performance degradation in deep learning-based segmentation models. Given the high cost of manual labeling and the need for privacy protection, it is often challenging to annotate the testing (target) domain data for model fine-tuning or to collect data from different domains to train domain generalization models. Therefore, using only unlabeled target domain data for test-time adaptation (TTA) presents a more practical but challenging solution. To improve the segmentation accuracy of deep learning-based models on unseen datasets, and especially to enhance the efficiency and stability of TTA for individual samples from heterogeneous domains. In this study, we proposed to dynamically adapt a wavelet-VNet (WaVNet) to unseen target domains with a hybrid objective function, based on each unlabeled test sample during the test time. We embedded multiscale wavelet coefficients into a V-Net encoder and adaptively adjusted the spatial and spectral features according to the input, and the model parameters were optimized by three loss functions. We integrated a shape-aware loss to focus on the foreground segmentations, a Refine loss to correct the incomplete and noisy segmentations caused by domain shifts, and an entropy loss to promote the global consistency of the segmentations. We evaluated the proposed method on multidomain liver and prostate segmentation datasets to assess its advantages over other TTA methods. For the source domain model training of the liver dataset, we used 15 3D MR image samples for training and 5 for validation. Correspondingly, for the prostate dataset, we used 22 3D MR image samples for training and 7 for validation. In the target domain, we used a single 3D MR image sample for adaptation and testing. The total number of testing samples is 60 in the liver dataset (for 3 different domains) and 116 in the prostate dataset (for 6 different domains). The proposed method showed the highest segmentation accuracy among all methods, achieving a mean(±SD) Dice coefficient (DSC) of 78.10±5.23% and a mean 95th Hausdorff distance (HD95) of 15.52±5.84mm on the liver dataset; and a mean DSC of 80.02±3.89% and a mean HD95 of 9.18±3.47mm on the prostate dataset. The DSC is 11.67% (in absolute terms) and 15.27% higher than that of the baseline (no adaptation) method, for the liver and the prostate datasets, respectively. The proposed adaptive WaVNet enhanced the image segmentation accuracy from unseen domains during the test time via unsupervised learning and multi-objective optimization. It can benefit clinical applications where data are scarce or with changing data distributions, including online adaptive radiotherapy. The code will be released at: https://github.com/sanny1226/WaVNet.

USMDA: Unsupervised Multisource Domain Adaptive ADHD prediction model using neuroimaging

There is an increasing number of large-scale cross-site database collections of neuroimaging markers (sMRI and fMRI) for studying neurodevelopmental illnesses (NDDs). Although a huge amount of data favors machine learning-based categorization algorithms, the unique heterogeneity of each site can impair cross-site generalization capacity. It is critical to create Unsupervised domain adaption methods for NDDs because obtaining appropriate diagnoses or labeling for NDDs might be problematic. In our work, we focus on Attention-deficit/hyperactivity disorder, which is the most common and frequently co-occurring NDD. We present an unsupervised multisource domain adaptation network (USMDA) with four primary components: Domain Alignment Module, Discrepancy Estimator, Pre-trained Model Generator, and Unsupervised Network. The Domain Alignment module is intended to incrementally and effectively align graph representations of the source and target domains. At the same time, the binary cross entropy regularizer is introduced for the first time during the training of a model learned on multiple source domains to improve existing feature alignment methods such as Transfer Joint Matching (TJM) and Joint Distribution Adaptation (JDA) by learning good unsupervised features. In an unsupervised network, the grid search optimization technique generates the optimal pseudo labels for unlabeled target data. We validate our proposed technique first on existing feature-level DA methods such as JDA, TJM, and Correlation alignment (CORAL), on the publically accessible dataset ADHD-200 and then by using binary cross-entropy in existing DA methods such as TJMCE and JDACE. The experimental results show that our proposed USMDAJDACE method when applied to multisite sMRI and fMRI ADHD data, can significantly outperform competitive methods for multi-center ADHD diagnosis.

Anti-forgetting source-free domain adaptation method for machine fault diagnosis

Unsupervised domain adaptation methods have made significant progress in the field of machine fault diagnosis. However, these methods generally assume the accessibility of source domain data during the cross-domain transfer phase. In practice, this assumption is often impractical due to data privacy requirements and the burden of data transmission and storage. Additionally, a key challenge lies in preventing the forgetting of source domain knowledge while performing cross-domain diagnosis in a source-free scenario. To address these issues, this paper proposes an anti-forgetting source-free domain adaptation method for machine fault diagnosis. Specifically, a prototype-based pseudo-label generation strategy, combined with a nearest-neighbor constraint, is employed to deeply explore diagnosis information in the unlabeled target domain data. Then, a fast nuclear-norm maximization constraint is introduced to reduce the density of target domain samples near the classification boundary, thereby decreasing the probability of misdiagnosis. Furthermore, a weighted Fisher regularization is designed to retain the diagnosis information of healthy states inherent in the source domain, thus mitigating information forgetting. Finally, comprehensive experiments are conducted to validate the superiority of our method in source-free cross-domain diagnosis from multiple perspectives, while also demonstrating its anti-forgetting properties.

Logical Relation Inference and Multiview Information Interaction for Domain Adaptation Person Re-Identification.

Domain adaptation person re-identification (Re-ID) is a challenging task, which aims to transfer the knowledge learned from the labeled source domain to the unlabeled target domain. Recently, some clustering-based domain adaptation Re-ID methods have achieved great success. However, these methods ignore the inferior influence on pseudo-label prediction due to the different camera styles. The reliability of the pseudo-label plays a key role in domain adaptation Re-ID, while the different camera styles bring great challenges for pseudo-label prediction. To this end, a novel method is proposed, which bridges the gap of different cameras and extracts more discriminative features from an image. Specifically, an intra-to-intermechanism is introduced, in which samples from their own cameras are first grouped and then aligned at the class level across different cameras followed by our logical relation inference (LRI). Thanks to these strategies, the logical relationship between simple classes and hard classes is justified, preventing sample loss caused by discarding the hard samples. Furthermore, we also present a multiview information interaction (MvII) module that takes features of different images from the same pedestrian as patch tokens, obtaining the global consistency of a pedestrian that contributes to the discriminative feature extraction. Unlike the existing clustering-based methods, our method employs a two-stage framework that generates reliable pseudo-labels from the views of the intracamera and intercamera, respectively, to differentiate the camera styles, subsequently increasing its robustness. Extensive experiments on several benchmark datasets show that the proposed method outperforms a wide range of state-of-the-art methods. The source code has been released at https://github.com/lhf12278/LRIMV.

BiPC: Bidirectional Probability Calibration for Unsupervised Domain Adaption

Unsupervised Domain Adaptation (UDA) leverages a labeled source domain to solve tasks in an unlabeled target domain. While Transformer-based methods have shown promise in UDA, their application is limited to plain Transformers, excluding Convolutional Neural Networks (CNNs) and hierarchical Transformers. To address this issues, we propose Bidirectional Probability Calibration (BiPC) from a probability space perspective. We demonstrate that the probability outputs from a pre-trained head, after extensive pre-training, are robust against domain gaps and can adjust the probability distribution of the task head. Moreover, the task head can enhance the pre-trained head during adaptation training, improving model performance through bidirectional complementation. Technically, we introduce Calibrated Probability Alignment (CPA) to adjust the pre-trained head’s probabilities, such as those from an ImageNet-1k pre-trained classifier. Additionally, we design a Calibrated Gini Impurity (CGI) loss to refine the task head, with calibrated coefficients learned from the pre-trained classifier. BiPC is a simple yet effective method applicable to various networks, including CNNs and Transformers. Experimental results demonstrate its remarkable performance across multiple UDA tasks. Our code will be available at: https://github.com/Wenlve-Zhou/BiPC.

FireDA: A Domain Adaptation-Based Method for Forest Fire Recognition with Limited Labeled Scenarios

Vision-based forest fire detection systems have significantly advanced through Deep Learning (DL) applications. However, DL-based models typically require large-scale labeled datasets for effective training, where the quality of data annotation is crucial to their performance. To address challenges related to the quality and quantity of labeling, a domain adaptation-based approach called FireDA is proposed for forest fire recognition in scenarios with limited labels. Domain adaptation, a subfield of transfer learning, facilitates the transfer of knowledge from a labeled source domain to an unlabeled target domain. The construction of the source domain FBD is initiated, which includes three common fire scenarios: forest (F), brightness (B), and darkness (D), utilizing publicly available labeled data. Subsequently, a novel algorithm called Neighborhood Aggregation-based 2-Stage Domain Adaptation (NA2SDA) is proposed. This method integrates feature distribution alignment with target domain Proxy Classification Loss (PCL), leveraging a neighborhood aggregation mechanism and a memory bank designed for the unlabeled samples in the target domain. This mechanism calibrates the source classifier and generates more accurate pseudo-labels for the unlabeled sample. Consequently, based on these pseudo-labels, the Local Maximum Mean Discrepancy (LMMD) and the Proxy Classification Loss (PCL) are computed. To validate the efficacy of the proposed method, the publicly available forest fire dataset, FLAME, is employed as the target domain for constructing a transfer learning task. The results demonstrate that our method achieves performance comparable to the supervised Convolutional Neural Network (CNN)-based state-of-the-art (SOTA) method, without requiring access to labels from the FLAME training set. Therefore, our study presents a viable solution for forest fire recognition in scenarios with limited labeling and establishes a high-accuracy benchmark for future research.

DCST: Dual Cross-Supervision for Transformer-based Unsupervised Domain Adaptation

Unsupervised Domain Adaptation aims to leverage a source domain with ample labeled data to tackle tasks on an unlabeled target domain. However, this poses a significant challenge, particularly in scenarios exhibiting significant disparities between the two domains. Prior methods often fall short in challenging domains due to the impact of incorrect pseudo-labeling noise and the limits of handcrafted domain alignment rules. In this paper, we propose a novel method called DCST (Dual Cross-Supervision Transformer), which improves upon existing methods in two key aspects. Firstly, vision transformer is combined with dual cross-supervision learning strategy to enforce consistency learning from different domains. The network accomplishes domain-specific self-training and cross-domain feature alignment in an adaptive manner. Secondly, due to the presence of noise in challenging domain, and the need to reduce the risks of model collapse and overfitting, we propose a Domain Shift Filter. Specifically, this module allows the model to leverage the memory of source domain features to facilitate a smooth transition. It can also improve the effectiveness of knowledge transfer between domains with significant gaps. We conduct extensive experiments on four benchmark datasets and achieved the best classification results, including 94.3% on Office-31, 86.0% on Office-Home, 89.3% on VisDA-2017, and 48.8% on DomainNet. Code is available in https://github.com/Yislight/DCST.

A two-layer regression network for robust and accurate domain adaptation

Naturally, classification models suffer from severe performance degradation when they are tested on datasets different from the ones used in training. Unsupervised domain adaptation helps to improve the generalizability of a pre-trained model by transferring knowledge from the labeled source domain (i.e., training dataset) to the unlabeled target domain (i.e., test dataset). The typical way of domain adaptation is to align the data distributions in embedding spaces between source and target domains. However, most existing works only enhance cross-domain consistency in the latent space disregarding the impact of source samples on the target classification task. Besides, discovering a subspace shared by two domains first and then training a transfer classifier separately hardly ensures that the obtained target features are suited for the classification model. To address these issues, in this work, we propose a two-layer regression network for domain adaptation (TRN-DA). TRN-DA learns class-wise domain invariant feature representations by jointly optimizing three tasks: supervised classification of source domain, unsupervised reconstruction of target domain, and alignment of cross-domain distributions. Furthermore, we introduce the concept of weight (importance) for source instances so that the resulting classifier can adapt well to the target domain. We formulate the domain adaptation problem as a unified optimization problem and solve it in an iterative way. Experimental results on different cross-domain image classification tasks demonstrate the encouraging performance of our TRN-DA compared to several recent state-of-the-art methods.