Domain Shift Research Articles

Improved beamforming by accounting for multiple domain shifts and sources of degradation

A number of mechanisms are known to degrade ultrasound images. To address these issues, many beamforming strategies have been proposed including a slew of recent approaches centered around deep learning techniques. Most deep learning methods target a single type of image degradation. Here, we will simultaneously address multiple sources of image degradation within the same beamformer. In particular, we will address phase aberration, reverberation and off-axis scatting within a single network. Within this context of training networks to address multiple sources of image degradation, we will highlight the problem of domain shift in ultrasound beamforming deep networks. Domain shift is the mismatch in the statistics of the training data and the data encountered in the intended use case that reduces the performance of the trained deep network. We will show that ultrasound beamformers struggle with domain shift on both the input side and the output side of the networks, and we will present strategies for addressing both domain shifts including the shift on the output side where there is no ground truth data to learn from.

Cerebral asymmetry representation learning-based deep subdomain adaptation network for electroencephalogram-based emotion recognition.

Objective.Extracting discriminative spatial information from multiple electrodes is a crucial and challenging problem for electroencephalogram (EEG)-based emotion recognition. Additionally, the domain shift caused by the individual differences degrades the performance of cross-subject EEG classification.Approach.To deal with the above problems, we propose the cerebral asymmetry representation learning-based deep subdomain adaptation network (CARL-DSAN) to enhance cross-subject EEG-based emotion recognition. Specifically, the CARL module is inspired by the neuroscience findings that asymmetrical activations of the left and right brain hemispheres occur during cognitive and affective processes. In the CARL module, we introduce a novel two-step strategy for extracting discriminative features through intra-hemisphere spatial learning and asymmetry representation learning. Moreover, the transformer encoders within the CARL module can emphasize the contributive electrodes and electrode pairs. Subsequently, the DSAN module, known for its superior performance over global domain adaptation, is adopted to mitigate domain shift and further improve the cross-subject performance by aligning relevant subdomains that share the same class samples.Main Results.To validate the effectiveness of the CARL-DSAN, we conduct subject-independent experiments on the DEAP database, achieving accuracies of 68.67% and 67.11% for arousal and valence classification, respectively, and corresponding accuracies of 67.70% and 67.18% on the MAHNOB-HCI database.Significance.The results demonstrate that CARL-DSAN can achieve an outstanding cross-subject performance in both arousal and valence classification.

I2F: A Unified Image-to-Feature Approach for Domain Adaptive Semantic Segmentation.

Unsupervised domain adaptation (UDA) for semantic segmentation is a promising task freeing people from heavy annotation work. However, domain discrepancies in low-level image statistics and high-level contexts compromise the segmentation performance over the target domain. A key idea to tackle this problem is to perform both image-level and feature-level adaptation jointly. Unfortunately, there is a lack of such unified approaches for UDA tasks in the existing literature. This paper proposes a novel UDA pipeline for semantic segmentation that unifies image-level and feature-level adaptation. Concretely, for image-level domain shifts, we propose a global photometric alignment module and a global texture alignment module that align images in the source and target domains in terms of image-level properties. For feature-level domain shifts, we perform global manifold alignment by projecting pixel features from both domains onto the feature manifold of the source domain; and we further regularize category centers in the source domain through a category-oriented triplet loss, and perform target domain consistency regularization over augmented target domain images. Experimental results demonstrate that our pipeline significantly outperforms previous methods. In the commonly tested GTA5 →Cityscapes task, our proposed method using Deeplab V3+ as the backbone surpasses previous SOTA by 8%, achieving 58.2% in mIoU.

Bibimbap : Pre-trained models ensemble for Domain Generalization

This paper addresses a machine learning problem often challenged by differences in the distributions of training and real-world data. We propose a framework that addresses the problem of underfitting in the ensembling method using pre-trained models and improves the performance and robustness of deep learning models through ensemble diversity. For the naive weight ensembling framework, we discovered that the ensembled models could not lie in the same loss basin under extreme domain shift conditions, suggesting that a loss barrier may exist. We used a fine-tuning step after the weighted ensemble to address the underfitting problem caused by the loss barrier and stabilize the batch normalization running parameters. We also inferred through qualitative analysis that the diversity of ensemble models affects domain generalization. We validate our method on a large-scale image dataset (ImageNet-1K) and chemical molecule data, which is suitable for testing with domain shift problems due to its data-splitting method.

Generalization of deep learning models for hepatic steatosis grading using B-mode ultrasound images

Grayscale ultrasound remains a key modality for screening of hepatic steatosis due to its non-invasiveness and availability. While neural networks have shown promise in this field, their main drawback lies in their inability to generalize to diverse real-world settings. Variations in equipment, acquisition parameters, or population significantly affect model performance. Test-time adaptation, an unsupervised domain adaptation technique, overcomes these limitations by adjusting trained models during inference. Our retrospective study used two datasets collected in separate populations, with different scanners and protocols. We propose an adaptation method, using test-time batch normalization to selectively adjust BatchNorm layers based on test data for predicting steatosis grades. Comparing the non-adapted and adapted models, the mean absolute error (± standard deviation) in grading four severities of steatosis decreased from 0.92 ± 0.21 to 0.64 ± 0.22 . Specifically, for detection of steatosis the area under the curve increased from 0.76 ± 0.05 to 0.95 ± 0.02 when using the adapted model. Adapted models show promising results in improving performance compared to base models when testing data differ significantly from training data. Results suggest that the proposed method effectively addresses domain shift in diagnosing fatty liver using ultrasound images, reducing risks associated with deploying trained models.

Zero-Shot Learning With Attentive Region Embedding and Enhanced Semantics.

The performance of zero-shot learning (ZSL) can be improved progressively by learning better features and generating pseudosamples for unseen classes. Existing ZSL works typically learn feature extractors and generators independently, which may shift the unseen samples away from their real distribution and suffers from the domain bias problem. In this article, to tackle this challenge, we propose a variational autoencoder (VAE)-based framework, that is, joint Attentive Region Embedding with Enhanced Semantics (AREES), which is tailored to advance the zero-shot recognition. Specifically, AREES is end-to-end trainable and consists of three network branches: 1) attentive region embedding is used to learn the semantic-guided visual features by the attention mechanism (AM); 2) a decomposition structure and a semantic pivot regularization are used to extract enhanced semantics; and 3) a multimodal VAE (mVAE) with the cross-reconstruction loss and the distribution alignment loss is used to obtain a shared latent embedding space of visual features and semantics. Finally, features' extraction and features' generation are optimized together in AREES to address the domain shift problem to a large extent. The comprehensive evaluations on six benchmarks, including the ImageNet, demonstrate the superiority of the proposed model over its state-of-the-art counterparts.

Semantics-Guided Contrastive Network for Zero-Shot Object Detection.

Zero-shot object detection (ZSD), the task that extends conventional detection models to detecting objects from unseen categories, has emerged as a new challenge in computer vision. Most existing approaches tackle the ZSD task with a strict mapping-transfer strategy that may lead to suboptimal ZSD results: 1) the learning process of these models neglects the available semantic information on unseen classes, which can easily bias towards the seen categories; 2) the original visual feature space is not well-structured for the ZSD task due to the lack of discriminative information. To address these issues, we develop a novel Semantics-Guided Contrastive Network for ZSD, named ContrastZSD, a detection framework that first brings contrastive learning mechanism into the realm of zero-shot detection. Particularly, ContrastZSD incorporates two semantics-guided contrastive learning subnets that contrast between region-category and region-region pairs respectively. The pairwise contrastive tasks take advantage of supervision signals derived from both the ground truth label and class similarity information. By performing supervised contrastive learning over those explicit semantic supervision, the model can learn more knowledge about unseen categories to avoid the bias problem to seen concepts, while optimizing the visual data structure to be more discriminative for better visual-semantic alignment. Extensive experiments are conducted on two popular benchmarks for ZSD, i.e., PASCAL VOC and MS COCO. Results show that our method outperforms the previous state-of-the-art on both ZSD and generalized ZSD tasks.

Two-stage adversarial learning based unsupervised domain adaptation for retinal OCT segmentation.

Deep learning based optical coherence tomography (OCT) segmentation methods have achieved excellent results, allowing quantitative analysis of large-scale data. However, OCT images are often acquired by different devices or under different imaging protocols, which leads to serious domain shift problem. This in turn results in performance degradation of segmentation models. Aiming at the domain shift problem, we propose a two-stage adversarial learning based network (TSANet) that accomplishes unsupervised cross-domain OCT segmentation. In the first stage, a Fourier transform based approach is adopted to reduce image style differences from the image level. Then, adversarial learning networks, including a segmenter and a discriminator, are designed to achieve inter-domain consistency in the segmentation output. In the second stage, pseudo labels of selected unlabeled target domain training data are used to fine-tune the segmenter, which further improves its generalization capability. The proposed method was tested on cross-domain datasets for choroid or retinoschisis segmentation tasks. For choroid segmentation, the model was trained on 400 images and validated on 100 images from the source domain, and then trained on 1320 unlabeled images and tested on 330 images from target domain I, and also trained on 400 unlabeled images and tested on 200 images from target domain II. For retinoschisis segmentation, the model was trained on 1284 images and validated on 312 images from the source domain, and then trained on 1024 unlabeled images and tested on 200 images from the target domain. The proposed method achieved significantly improved results over that without domain adaptation, with improvement of 8.34%, 55.82% and 3.53% in intersection over union (IoU) respectively for the three test sets. The performance is better than some state-of-the-art domain adaptation methods. The proposed TSANet, with image level adaptation, feature level adaptation and pseudo-label based fine-tuning, achieved excellent cross-domain generalization. This alleviates the burden of obtaining additional manual labels when adapting the deep learning model to new OCT data.

Privacy-Preserving Breast Cancer Classification: A Federated Transfer Learning Approach.

Breast cancer is deadly cancer causing a considerable number of fatalities among women in worldwide. To enhance patient outcomes as well as survival rates, early and accurate detection is crucial. Machine learning techniques, particularly deep learning, have demonstrated impressive success in various image recognition tasks, including breast cancer classification. However, the reliance on large labeled datasets poses challenges in the medical domain due to privacy issues and data silos. This study proposes a novel transfer learning approach integrated into a federated learning framework to solve the limitations of limited labeled data and data privacy in collaborative healthcare settings. For breast cancer classification, the mammography and MRO images were gathered from three different medical centers. Federated learning, an emerging privacy-preserving paradigm, empowers multiple medical institutions to jointly train the global model while maintaining data decentralization. Our proposed methodology capitalizes on the power of pre-trained ResNet, a deep neural network architecture, as a feature extractor. By fine-tuning the higher layers of ResNet using breast cancer datasets from diverse medical centers, we enable the model to learn specialized features relevant to different domains while leveraging the comprehensive image representations acquired from large-scale datasets like ImageNet. To overcome domain shift challenges caused by variations in data distributions across medical centers, we introduce domain adversarial training. The model learns to minimize the domain discrepancy while maximizing classification accuracy, facilitating the acquisition of domain-invariant features. We conducted extensive experiments on diverse breast cancer datasets obtained from multiple medical centers. Comparative analysis was performed to evaluate the proposed approach against traditional standalone training and federated learning without domain adaptation. When compared with traditional models, our proposed model showed a classification accuracy of 98.8% and a computational time of 12.22s. The results showcase promising enhancements in classification accuracy and model generalization, underscoring the potential of our method in improving breast cancer classification performance while upholding data privacy in a federated healthcare environment.

TMNIO:Triplet merged network with involution operators for improved few‐shot image classification

AbstractFew‐shot learning enables machines to learn efficiently from limited labelled data. However, existing few‐shot learning methods may perform poorly when there is a lack of sufficient samples, and may encounter problems such as domain shift or overfitting when applied to new domains or tasks. To address the issues of poor fitting and insufficient generalization ability in new domains, a new method called triplet merged network with involution operators (TMNIO) is proposed. This method employs dual encoders that extract common and distinctive features from the prototype network, thereby enhancing the model's feature extraction capability. To further improve this ability, the traditional convolutional kernels are replaced with involution operators, which not only reduce the parameter count but also enlarge the receptive field to better extract local feature information. Additionally, this method employs a two‐stage training strategy, where triplet loss is used in the first stage to train the model and enhance its robustness and generalization ability. Extensive experiments on the miniImageNet, Omniglot, and Caltech‐UCSD Birds‐200 (CUB) datasets have shown that our proposed method achieved significant improvements in both training speed and accuracy, particularly on the miniImageNet dataset, where it achieved an outstanding 10% performance improvement.

Deep learning for head and neck semi-supervised semantic segmentation

Objective. Radiation therapy (RT) represents a prevalent therapeutic modality for head and neck (H&N) cancer. A crucial phase in RT planning involves the precise delineation of organs-at-risks (OARs), employing computed tomography (CT) scans. Nevertheless, the manual delineation of OARs is a labor-intensive process, necessitating individual scrutiny of each CT image slice, not to mention that a standard CT scan comprises hundreds of such slices. Furthermore, there is a significant domain shift between different institutions’ H&N data, which makes traditional semi-supervised learning strategies susceptible to confirmation bias. Therefore, effectively using unlabeled datasets to support annotated datasets for model training has become a critical issue for preventing domain shift and confirmation bias. Approach. In this work, we proposed an innovative cross-domain orthogon-based-perspective consistency (CD-OPC) strategy within a two-branch collaborative training framework, which compels the two sub-networks to acquire valuable features from unrelated perspectives. More specifically, a novel generative pretext task cross-domain prediction (CDP) was designed for learning inherent properties of CT images. Then this prior knowledge was utilized to promote the independent learning of distinct features by the two sub-networks from identical inputs, thereby enhancing the perceptual capabilities of the sub-networks through orthogon-based pseudo-labeling knowledge transfer. Main results. Our CD-OPC model was trained on H&N datasets from nine different institutions, and validated on the four local intuitions’ H&N datasets. Among all datasets CD-OPC achieved more advanced performance than other semi-supervised semantic segmentation algorithms. Significance. The CD-OPC method successfully mitigates domain shift and prevents network collapse. In addition, it enhances the network’s perceptual abilities, and generates more reliable predictions, thereby further addressing the confirmation bias issue.

TFFNet: A robust approach with anti-noise and domain shift adaptation for intelligent fault diagnosis of rotating machinery

Recently, deep learning has been a predominantly used technique for intelligent fault diagnosis of industrial machines. It has accomplished satisfactory performance as well. However, noise is present in a real-life industrial working environment, and the operational load also constantly changes. This work proposes a Time-Frequency Fusion Network (TFFNet) for intelligent fault diagnosis. It is robust convolutional neural network based deep-learning algorithm and eliminates the signal processing required for denoising. The success of the developed model is verified in the presence of real-time noisy conditions and under a load-varying environment. The proposed model attained 99.98% accuracy in a noisy environment and 98.6% average accuracy under six cases of domain shift. Finally, the results are compared with past studies using accuracy as a performance indicator.

Exploring biometric domain adaptation in human action recognition models for unconstrained environments

AbstractIn conventional machine learning (ML), a fundamental assumption is that the training and test sets share identical feature distributions, a reasonable premise drawn from the same dataset. However, real-world scenarios often defy this assumption, as data may originate from diverse sources, causing disparities between training and test data distributions. This leads to a domain shift, where variations emerge between the source and target domains. This study delves into human action recognition (HAR) models within an unconstrained, real-world setting, scrutinizing the impact of input data variations related to contextual information and video encoding. The objective is to highlight the intricacies of model performance and interpretability in this context. Additionally, the study explores the domain adaptability of HAR models, specifically focusing on their potential for re-identifying individuals within uncontrolled environments. The experiments involve seven pre-trained backbone models and introduce a novel analytical approach by linking domain-related (HAR) and domain-unrelated (re-identification (re-ID)) tasks. Two key analyses addressing contextual information and encoding strategies reveal that maintaining the same encoding approach during training results in high task correlation while incorporating richer contextual information enhances performance. A notable outcome of this study is the comprehensive evaluation of a novel transformer-based architecture driven by a HAR backbone, which achieves a robust re-ID performance superior to state-of-the-art (SOTA). However, it faces challenges when other encoding schemes are applied, highlighting the role of the HAR classifier in performance variations.

MISNet: multi-source information-shared EEG emotion recognition network with two-stream structure.

When constructing machine learning and deep neural networks, the domain shift problem on different subjects complicates the subject independent electroencephalography (EEG) emotion recognition. Most of the existing domain adaptation methods either treat all source domains as equivalent or train source-specific learners directly, misleading the network to acquire unreasonable transfer knowledge and thus resulting in negative transfer. This paper incorporates the individual difference and group commonality of distinct domains and proposes a multi-source information-shared network (MISNet) to enhance the performance of subject independent EEG emotion recognition models. The network stability is enhanced by employing a two-stream training structure with loop iteration strategy to alleviate outlier sources confusing the model. Additionally, we design two auxiliary loss functions for aligning the marginal distributions of domain-specific and domain shared features, and then optimize the convergence process by constraining gradient penalty on these auxiliary loss functions. Furthermore, the pre-training strategy is also proposed to ensure that the initial mapping of shared encoder contains sufficient emotional information. We evaluate the proposed MISNet to ascertain the impact of several hyper-parameters on the domain adaptation capability of network. The ablation experiments are conducted on two publically accessible datasets SEED and SEED-IV to assess the effectiveness of each loss function. The experimental results demonstrate that by disentangling private and shared emotional characteristics from differential entropy features of EEG signals, the proposed MISNet can gain robust subject independent performance and strong domain adaptability.

Fusing consensus knowledge: A federated learning method for fault diagnosis via privacy-preserving reference under domain shift

Recently, federated fault diagnosis has garnered growing attention due to its promising capabilities in information fusion with privacy preservation. However, most of the existing approaches are based on the assumptions of no domain shift between multiple factories and no unseen domains for online applications. In actual industry, these assumptions are generally unsatisfied due to prominent environmental noises, mechanical wear, and changes in working conditions. Federated models that ignore domain shifts would face the negative aggregation problem and are not robust to unseen domains. To solve the domain shift problem, a federated domain generalization method is proposed for privacy-preserving fault diagnosis in this article. The key idea is to construct a sharable reference domain in cloud, which can convert the privacy-risky centralized alignment problem into a privacy-preserving pairwise alignment problem. Based on the recognition that any fault category in a discriminative feature space can be characterized by a particular position and volatility, we design a shareable domain generator to provide a reference for pairwise alignment. Then, the non-deterministic sampling and non-parametric alignment criterion are introduced to realize local domain alignment, which facilitates the domain-invariant feature extraction. Finally, by the alternation of local domain alignment and global reference synchronization, the alignment of multi-source domains is achieved implicitly. We give convergence guarantees for the proposed method and derive the generalization error bound of federated DG, which illustrates the positive effect of the proposed method on improving generalization. Experiments on two cases demonstrate the consistent superior generalization performance of our method without the risk of data leakage.

Relative Norm Alignment for Tackling Domain Shift in Deep Multi-modal Classification

Multi-modal learning has gained significant attention due to its ability to enhance machine learning algorithms. However, it brings challenges related to modality heterogeneity and domain shift. In this work, we address these challenges by proposing a new approach called Relative Norm Alignment (RNA) loss. RNA loss exploits the observation that variations in marginal distributions between modalities manifest as discrepancies in their mean feature norms, and rebalances feature norms across domains, modalities, and classes. This rebalancing improves the accuracy of models on test data from unseen (“target”) distributions. In the context of Unsupervised Domain Adaptation (UDA), we use unlabeled target data to enhance feature transferability. We achieve this by combining RNA loss with an adversarial domain loss and an Information Maximization term that regularizes predictions on target data. We present a comprehensive analysis and ablation of our method for both Domain Generalization and UDA settings, testing our approach on different modalities for tasks such as first and third person action recognition, object recognition, and fatigue detection. Experimental results show that our approach achieves competitive or state-of-the-art performance on the proposed benchmarks, showing the versatility and effectiveness of our method in a wide range of applications.

ST-GAN: A Swin Transformer-Based Generative Adversarial Network for Unsupervised Domain Adaptation of Cross-Modality Cardiac Segmentation.

Unsupervised domain adaptation (UDA) methods have shown great potential in cross-modality medical image segmentation tasks, where target domain labels are unavailable. However, the domain shift among different image modalities remains challenging, because the conventional UDA methods are based on convolutional neural networks (CNNs), which tend to focus on the texture of images and cannot establish the global semantic relevance of features due to the locality of CNNs. This paper proposes a novel end-to-end Swin Transformer-based generative adversarial network (ST-GAN) for cross-modality cardiac segmentation. In the generator of ST-GAN, we utilize the local receptive fields of CNNs to capture spatial information and introduce the Swin Transformer to extract global semantic information, which enables the generator to better extract the domain-invariant features in UDA tasks. In addition, we design a multi-scale feature fuser to sufficiently fuse the features acquired at different stages and improve the robustness of the UDA network. We extensively evaluated our method with two cross-modality cardiac segmentation tasks on the MS-CMR 2019 dataset and the M&Ms dataset. The results of two different tasks show the validity of ST-GAN compared with the state-of-the-art cross-modality cardiac image segmentation methods.

Named entity recognition using transfer learning and small human‐ and meta‐pseudo‐labeled datasets

AbstractWe introduce a high‐performance named entity recognition (NER) model for written and spoken language. To overcome challenges related to labeled data scarcity and domain shifts, we use transfer learning to leverage our previously developed KorBERT as the base model. We also adopt a meta‐pseudo‐label method using a teacher/student framework with labeled and unlabeled data. Our model presents two modifications. First, the student model is updated with an average loss from both human‐ and pseudo‐labeled data. Second, the influence of noisy pseudo‐labeled data is mitigated by considering feedback scores and updating the teacher model only when below a threshold (0.0005). We achieve the target NER performance in the spoken language domain and improve that in the written language domain by proposing a straightforward rollback method that reverts to the best model based on scarce human‐labeled data. Further improvement is achieved by adjusting the label vector weights in the named entity dictionary.

Non-Fluent Synthetic Target-Language Data Improve Neural Machine Translation.

When the amount of parallel sentences available to train a neural machine translation is scarce, a common practice is to generate new synthetic training samples from them. A number of approaches have been proposed to produce synthetic parallel sentences that are similar to those in the parallel data available. These approaches work under the assumption that non-fluent target-side synthetic training samples can be harmful and may deteriorate translation performance. Even so, in this paper we demonstrate that synthetic training samples with non-fluent target sentences can improve translation performance if they are used in a multilingual machine translation framework as if they were sentences in another language. We conducted experiments on ten low-resource and four high-resource translation tasks and found out that this simple approach consistently improves translation performance as compared to state-of-the-art methods for generating synthetic training samples similar to those found in corpora. Furthermore, this improvement is independent of the size of the original training corpus, the resulting systems are much more robust against domain shift and produce less hallucinations.

Graph Information Aggregation Cross-Domain Few-Shot Learning for Hyperspectral Image Classification.

Most domain adaptation (DA) methods in cross-scene hyperspectral image classification focus on cases where source data (SD) and target data (TD) with the same classes are obtained by the same sensor. However, the classification performance is significantly reduced when there are new classes in TD. In addition, domain alignment, as one of the main approaches in DA, is carried out based on local spatial information, rarely taking into account nonlocal spatial information (nonlocal relationships) with strong correspondence. A graph information aggregation cross-domain few-shot learning (Gia-CFSL) framework is proposed, intending to make up for the above-mentioned shortcomings by combining FSL with domain alignment based on graph information aggregation. SD with all label samples and TD with a few label samples are implemented for FSL episodic training. Meanwhile, intradomain distribution extraction block (IDE-block) and cross-domain similarity aware block (CSA-block) are designed. The IDE-block is used to characterize and aggregate the intradomain nonlocal relationships and the interdomain feature and distribution similarities are captured in the CSA-block. Furthermore, feature-level and distribution-level cross-domain graph alignments are used to mitigate the impact of domain shift on FSL. Experimental results on three public HSI datasets demonstrate the superiority of the proposed method. The codes will be available from the website: https://github.com/YuxiangZhang-BIT/IEEE_TNNLS_Gia-CFSL.