Unsupervised Domain Adaptation Algorithm Research Articles

A Copula Based Unsupervised Domain Adaptation for Image Classification

In this paper, we present an unsupervised domain adaptation algorithm for image classification using principal component analysis (PCA) and Gaussian copula function alignment. The motivation of the proposed algorithm stems from the idea of CORAL algorithm which extracts domain invariant features by aligning the correlation structure between a source and a target domain. However, it suffers from the fact that highly skewed marginal distributions happen to distort the correlation structure so that it may cause a negative transfer. Therefore we utilize a copula function that enables us to analyze separately the dependency structure and the marginals by Sklar’s theorem. In particular, we propose to align the Gaussian copula correlations in the copula feature spaces instead of aligning the correlation matrices in the original space. Considering the extremely skewed distribution of SURF image features in Office-Caltech10 data set we apply PCA first in order to extract some skewness-mitigated principal features and then derive copula features to align for domain adaptation by using CORAL idea with Gaussian copula correlation matrices. The proposed method showed a good classification accuracy when applied to image classification problem in an unsupervised domain adaptation setting.

DANNMCTG: Domain-Adversarial Training of Neural Network for multicenter antenatal cardiotocography signal classification

Intelligent classification of cardiotocography (CTG) based on machine learning (ML), a useful tool to improve the accuracy of fetal abnormality detection, can assist obstetricians with clinical decisions. With the advancement of information technologies and medical devices, there are development opportunities for multicenter clinical research and obtaining more digital CTG signals. However, most of the existing clinical multicenter CTG datasets are partially annotated and have discrepancies which do not satisfy the ML condition of independent identical distribution. Therefore, this paper focuses on an unsupervised domain adaptation (UDA) algorithm to realize cross-domain intelligent classification of multimodal CTG signals. We propose a method dubbed domain adversarial training of neural network for multicenter CTG (DANNMCTG), which mainly consists of a label classifier, a feature extractor and a domain discriminator. To match different distribution of fetal heart rate (FHR), uterine contraction (UC) and fetal movement (FetMov) signals, we condition the domain alignment on label predictions by defining the multi-linear map. For analysis, two datasets from the hospital central station and home monitoring devices were considered as the source and target domains. The results showed that the accuracy value, F1 value and area under the curve (AUC) value of the DANNMCTG were 71.25%, 76.08% and 0.7705, respectively. This method significantly improved the performance of the deep learning models without exploiting any information in the target domain, and outperformed the state-of-the-art UDA algorithms for CTG classification. In summary, the DANNMCTG can effectively mitigate the influence of domain shift for multicenter intelligent prenatal fetal monitoring.

MGFKD: A semi-supervised multi-source domain adaptation algorithm for cross-subject EEG emotion recognition

Currently, most models rarely consider the negative transfer problem in the research field of cross-subject EEG emotion recognition. To solve this problem, this paper proposes a semi-supervised domain adaptive algorithm based on few labeled samples of target subject, which called multi-domain geodesic flow kernel dynamic distribution alignment (MGFKD). It consists of three modules: 1) GFK common feature extractor: projects the feature distribution of source and target subjects to the Grassmann manifold space, and obtains the latent common features of the two feature distributions through GFK method. 2) Source domain selector: obtains pseudo-labels of the target subject through weak classifier, finds "golden source subjects" by using few known labels of target subjects. 3) Label corrector: uses a dynamic distribution balance strategy to correct the pseudo-labels of the target subject. We conducted comparison experiments on the SEED and SEED-IV datasets, and the results show that MGFKD outperforms unsupervised and semi-supervised domain adaptation algorithms, achieving an average accuracy of 87.51±7.68% and 68.79±8.25% on the SEED and SEED-IV datasets with only one labeled sample per video for target subject. Especially when the number of source domains is set as 6 and the number of known labels is set as 5, the accuracy increase to 90.20±7.57% and 69.99±7.38%, respectively. The above results prove that our proposed algorithm can efficiently and quickly improve the cross-subject EEG emotion classification performance. Since it only need a small number of labeled samples of new subjects, making it has strong application value in future EEG-based emotion recognition applications.

Select, Purify, and Exchange: A Multisource Unsupervised Domain Adaptation Method for Building Extraction.

Accurately extracting buildings from aerial images has essential research significance for timely understanding human intervention on the land. The distribution discrepancies between diversified unlabeled remote sensing images (changes in imaging sensor, location, and environment) and labeled historical images significantly degrade the generalization performance of deep learning algorithms. Unsupervised domain adaptation (UDA) algorithms have recently been proposed to eliminate the distribution discrepancies without re-annotating training data for new domains. Nevertheless, due to the limited information provided by a single-source domain, single-source UDA (SSUDA) is not an optimal choice when multitemporal and multiregion remote sensing images are available. We propose a multisource UDA (MSUDA) framework SPENet for building extraction, aiming at selecting, purifying, and exchanging information from multisource domains to better adapt the model to the target domain. Specifically, the framework effectively utilizes richer knowledge by extracting target-relevant information from multiple-source domains, purifying target domain information with low-level features of buildings, and exchanging target domain information in an interactive learning manner. Extensive experiments and ablation studies constructed on 12 city datasets prove the effectiveness of our method against existing state-of-the-art methods, e.g., our method achieves 59.1% intersection over union (IoU) on Austin and Kitsap → Potsdam, which surpasses the target domain supervised method by 2.2% . The code is available at https://github.com/QZangXDU/SPENet.

Visualization of concrete internal defects based on unsupervised domain adaptation algorithm for transfer learning of experiment-simulation hybrid dataset of impact echo signals

Detecting concrete internal defects through deep learning analysis of impact echo signals faces two challenges: (1) the traditional signal processing method such as wavelet transform (WT) fails to reflect data-sensitive damage characteristics due to the uncertainty principle and (2) the limited labeled data acquired from real structures impedes network training. To address the first challenge, this paper proposes the WT-based synchrosqueezing transform (WT-SST) for the conversion of time-series data to the time-frequency spectrogram, which can provide effective features for the network in time and frequency domains simultaneously. To overcome the second challenge, numerical simulation data are supplemented for the augment of labeled data. To minimize the effect of data variance between experiments and simulations, this paper uses an unsupervised domain adaptation (DA) network for the transfer training of labeled simulation data (original domain) and unlabeled experimental data (target domain). The DA network extracts domain-invariant features by maximizing the domain recognition error and minimizing the probability distribution distance. The damage probability is calculated by the trained model to produce a 2D defect contour image of concrete specimens, and the three-dimensional visualization of internal defects by estimating the defect depth based on the defect area of contour image. Finally, the recognition precision, recall, F1-score, and accuracy of the model of unsupervised DA network trained by a hybrid dataset reaches 89.4%, 88.4%, 88.9%, and 94.7%, respectively.

A Domain-Adversarial Multi-Graph Convolutional Network for Unsupervised Domain Adaptation Rolling Bearing Fault Diagnosis

The transfer learning method, based on unsupervised domain adaptation (UDA), has been broadly utilized in research on fault diagnosis under variable working conditions with certain results. However, traditional UDA methods pay more attention to extracting information for the class labels and domain labels of data, ignoring the influence of data structure information on the extracted features. Therefore, we propose a domain-adversarial multi-graph convolutional network (DAMGCN) for UDA. A multi-graph convolutional network (MGCN), integrating three graph convolutional layers (multi-receptive field graph convolutional (MRFConv) layer, local extreme value convolutional (LEConv) layer, and graph attention convolutional (GATConv) layer) was used to mine data structure information. The domain discriminators and classifiers were utilized to model domain labels and class labels, respectively, and align the data structure differences through the correlation alignment (CORAL) index. The classification and feature extraction ability of the DAMGCN was significantly enhanced compared with other UDA algorithms by two example validation results, which can effectively achieve rolling bearing cross-domain fault diagnosis.

CSF: Closed-mask-guided semantic fusion method for semantic perception of unknown scenes

• A mask-guided semantic fusion method for optimizing the results of semantic segmentation. • A closed mask generation (CEG) module for post-processing of the edge detection algorithm. • A semantic confidence fusion (SCF) module to fuse the semantic segmentation results. • This method does not need to obtain the target data in advance. • This method has been verified in two scenarios. Though many high-precision semantic segmentation models have been proposed, how to improve the generalization ability of these models is still an urgent problem. Recently, a great number of unsupervised domain adaptation (UDA) algorithms around domain adaptation problems have been studied in semantic segmentation. These methods require labeled source domain data and unlabeled target domain data. In this paper, we propose a closed-mask-guided semantic fusion method (CSF) to improve the semantic segmentation results of unknown scenes, where the target domain data is not obtained in advance. First, a Closed Mask Generation (CMG) module is designed to convert the edge detection result into a mask that can segment the image into several image blocks. Then, a Semantic Confidence Fusion (SCF) module based on information entropy and voting method is introduced, which can select reliable semantic segmentation results for each image block by comparing the confidence of several semantic segmentation networks . In addition, the experimental results on both KITTI and COCO Stuff datasets validate the effectiveness of this method. The code is publicly available at https://github.com/tryhere/CSF .

Unsupervised multi-source domain adaptation with graph convolution network and multi-alignment in mixed latent space

This paper proposes an unsupervised multi-source domain adaptation algorithm with graph convolution network and multi-alignment in mixed latent space, which leverages domain labels, data structure, and category labels in a unified network but improves domain-invariant semantic representation by several innovations. Specifically, a novel data structure alignment is proposed to exploit the inherent properties of different domains while using current domain alignment and classification result alignment. Through this design, category consistency can be considered in both latent space, and domain and structure discrepancy between different source domains and the target domain can be eliminated. Moreover, we also use category alignment based on both CNN and GCN features to optimize category decision boundary. Experiment results show that the proposed method brings sufficient improvement especially for adaptation tasks with large shift in data distribution.

Evaluation of domain generalization and adaptation on improving model robustness to temporal dataset shift in clinical medicine

Temporal dataset shift associated with changes in healthcare over time is a barrier to deploying machine learning-based clinical decision support systems. Algorithms that learn robust models by estimating invariant properties across time periods for domain generalization (DG) and unsupervised domain adaptation (UDA) might be suitable to proactively mitigate dataset shift. The objective was to characterize the impact of temporal dataset shift on clinical prediction models and benchmark DG and UDA algorithms on improving model robustness. In this cohort study, intensive care unit patients from the MIMIC-IV database were categorized by year groups (2008–2010, 2011–2013, 2014–2016 and 2017–2019). Tasks were predicting mortality, long length of stay, sepsis and invasive ventilation. Feedforward neural networks were used as prediction models. The baseline experiment trained models using empirical risk minimization (ERM) on 2008–2010 (ERM[08–10]) and evaluated them on subsequent year groups. DG experiment trained models using algorithms that estimated invariant properties using 2008–2016 and evaluated them on 2017–2019. UDA experiment leveraged unlabelled samples from 2017 to 2019 for unsupervised distribution matching. DG and UDA models were compared to ERM[08–16] models trained using 2008–2016. Main performance measures were area-under-the-receiver-operating-characteristic curve (AUROC), area-under-the-precision-recall curve and absolute calibration error. Threshold-based metrics including false-positives and false-negatives were used to assess the clinical impact of temporal dataset shift and its mitigation strategies. In the baseline experiments, dataset shift was most evident for sepsis prediction (maximum AUROC drop, 0.090; 95% confidence interval (CI), 0.080–0.101). Considering a scenario of 100 consecutively admitted patients showed that ERM[08–10] applied to 2017–2019 was associated with one additional false-negative among 11 patients with sepsis, when compared to the model applied to 2008–2010. When compared with ERM[08–16], DG and UDA experiments failed to produce more robust models (range of AUROC difference, − 0.003 to 0.050). In conclusion, DG and UDA failed to produce more robust models compared to ERM in the setting of temporal dataset shift. Alternate approaches are required to preserve model performance over time in clinical medicine.

A Two-Way alignment approach for unsupervised multi-Source domain adaptation

Domain adaptation aims at transferring knowledge from labeled source domain to unlabeled target domain. Current advances primarily concern single source domain and neglect the setting of multiple source domains. Previous unsupervised multi-source domain adaptation (MDA) algorithms only consider domain-level alignment, while neglecting the category-level information among multiple domains and the instance variations inside each domain. This paper introduces a Two-Way alignment framework for MDA (TWMDA), which considers both domain-level and category-level alignments, and addresses the instance variations. We first align the target and multiple sources on the domain-level by an adversarial learning process. To circumvent the drawbacks of adversarial learning, we further reduce the domain gap on the category-level by minimizing the distance between the category prototypes and unlabeled target instances. To address the instance variations, we design an instance weighting strategy for diverse source instances. The effectiveness of TWMDA is demonstrated on three benchmark datasets for image classification.

M2N: Mutual constraint network for multi-level unsupervised domain adaptation

Unsupervised domain adaptation (UDA) algorithms aim to transfer the knowledge learned from the labeled source domain to the unlabeled target domain. To tackle this issue, domain-alignment models based on different representation space levels were used to extract knowledge relevant to the ultimate task. Although they achieved remarkable performance, domain shift that poses an evident challenge for adapting the model trained on one domain to another one was ignored. Pseudo labels learning for UDA is therefore proposed to reduce the impact of domain shifts. However, the inevitable label noise severely weakens the model’s capability of feature representations. Inspired by recent studies on deep learning with noisy labels, we propose a Mutual Constraint Network for multi-level UDA (M2N), to mitigate the effects of noisy pseudo labels and learn better feature representation from different space levels. Extensive experimental results on Digits, Office-31, and VisDA-2017, show that our methods can achieve new state-of-the-art performance on three benchmark tasks and improve significantly over prior single-level UDA.

Heterogeneous graph driven unsupervised domain adaptation of person re-identification

How to incrementally optimize a pre-trained classifier in an unlabeled target domain is a core challenging problem of domain adaptation (DA) for many visual tasks, such as Person Re-identification (re-ID). Most of the existing methods optimize the model based on pseudo labels or similarity of instance pairs, but ignoring the diverse manifold structures of unlabeled instances in the whole dataset. In this paper, we address the importance of such structural information in domain adaptation, and propose a Heterogeneous Graph driven Optimization scheme, namely H-GO, for structure based unsupervised learning. In particular, H-GO builds a heterogeneous graph of unlabeled images to consider the heterogeneous properties of images from various cameras with varied visual styles. A heterogeneous affinity propagation method is further applied to explore the graph based affinity between the instances which share similar manifold structures. Finally, a heterogeneous affinity learning procedure is taken to optimize the visual models by using the graph based affinity of instances. Comprehensive experiments are conducted on three large-scale re-ID datasets, and the results demonstrate the flexibility and the superior performance of H-GO than state-of-the-art unsupervised domain adaptation algorithms.

Abstract 242: Transferring diagnostic and prognostic molecular models across technological platforms

Abstract The reproducibility of results obtained using RNA data across labs is a major hurdle in cancer research. Difference in library preparation methods and gene expression quantification platforms prevent the application of trained models to new data across labs. SpinAdapt is a novel unsupervised domain adaptation algorithm that enables the transfer of existing molecular models across labs and technological platforms, without requiring re-training or calibration of existing models for future prospective data. Furthermore, SpinAdapt uses summary statistics (independent latent space representations) to calculate data corrections, rather than requiring full data access. This allows for transfer of molecular models across sequencing platforms and between labs without loss of data ownership or compromise of data privacy. To evaluate SpinAdapt, we performed two sets of experiments: A) We transferred molecular tumor subtype classifiers across four pairs of publicly available cancer datasets (bladder, breast, colorectal, pancreatic), covering 4,076 samples across 18 different tumor subtypes and three technological platforms (RNASeq, Affymetrix U133plus2, and HE1ST). For each pair of datasets we trained a subtype classifier on one dataset (target) according to well-accepted subtyping annotations (Zea Tan et al. 2019; Prat et al. 2012; Guinney et al. 2015; Bailey et al. 2016), and then evaluated the classifier accuracy on the other dataset (source). For each tumor subtype, we quantified the classification performance using mean AUC score across random subsets of the source dataset, where each subset was corrected using SpinAdapt. We aggregated performance across all subtypes and report the average mean AUC score for each cancer type: bladder 0.95, breast 0.98, colorectal 0.98, pancreatic 0.96; demonstrating high accuracy on all diagnostic tasks. B) To demonstrate the transferability of prognostic models, we trained five Cox survival models on five target cancer datasets respectively (breast, lung, colorectal, liver, pancreatic) ranging from 186 to 2,919 RNASeq samples. We used SpinAdapt to adapt five source cancer datasets to the target datasets, ranging from 226 to 1,038 samples across different platforms (RNASeq, Affymetrix U133Plus2 and HG-U133A, Illumina HumanHT-12v4). For every cancer type, we trained a Cox model on the target dataset, and measured its performance by predicting survival risk on the corresponding adapted source dataset. We show high survival prediction accuracy for all datasets (Log-rank P-values and c-index): lung [1e-6, 01.661], breast [5e-5, 0.626], liver [1e-4, 0.708], pancreatic [2e-4, 0.629], colorectal [9e-4, 0.661]. SpinAdapt transferred diagnostic and prognostic models over 14 cancer datasets covering 7,146 samples across six different cancer types and various platforms (RNASeq, microarray), while maintaining model accuracy and statistical significance. Citation Format: Talal Ahmed, Stephane Wenric, Mark Carty, Rafael Pelossof. Transferring diagnostic and prognostic molecular models across technological platforms [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 242.

Overview of Generative Adversarial Networks

The study of generative adversarial networks (GAN) provides a new approach and framework for computer vision and makes an outstanding choice for cross-domain image translation problems. In this work, the basic framework structure of GAN as well as the evolution of GAN are summarized. It is known that the unsupervised domain adaptation algorithms attempt to map representations between the two domains, or learn to extract features that are domain–invariant. An approach that learns in an unsupervised manner can be a transformation in the pixel space from one domain to the other. Finally, some existing problems of GAN and its potential future are also summarized and discussed.

Unsupervised Domain Adaptation Based on Correlation Maximization

This research proposes a novel unsupervised domain adaptation algorithm for cross-domain visual recognition. Distance Correlation-based Domain Adaptation or DCDA algorithm is developed by a correlation measure, called distance correlation. DCDA exploits both the statistical and geometrical properties of the data while embedding the both domain instances to a latent feature space. Unlike many proposed algorithms in the literature that utilize the source domain labels to learn pseudo labels, DCDA further exploits the available information in the source domain labels to discover an appropriate projection operator. The implementation of the proposed DCDA algorithm is easy, and it has a closed-form solution. Our experiments and analyses of the results over a wide variety of benchmark domain adaptation data sets indicate that DCDA has significantly better results in comparison with other state-of-the-art approaches in unsupervised domain adaptation and deep learning literature.

Incremental Real-Time Personalization in Human Activity Recognition Using Domain Adaptive Batch Normalization

Human Activity Recognition (HAR) from devices like smartphone accelerometers is a fundamental problem in ubiquitous computing. Machine learning based recognition models often perform poorly when applied to new users that were not part of the training data. Previous work has addressed this challenge by personalizing general recognition models to the unique motion pattern of a new user in a static batch setting. They require target user data to be available upfront. The more challenging online setting has received less attention. No samples from the target user are available in advance, but they arrive sequentially. Additionally, the motion pattern of users may change over time. Thus, adapting to new and forgetting old information must be traded off. Finally, the target user should not have to do any work to use the recognition system by, say, labeling any activities. Our work addresses all of these challenges by proposing an unsupervised online domain adaptation algorithm. Both classification and personalization happen continuously and incrementally in real time. Our solution works by aligning the feature distributions of all subjects, be they sources or the target, in hidden neural network layers. To this end, we normalize the input of a layer with user-specific mean and variance statistics. During training, these statistics are computed over user-specific batches. In the online phase, they are estimated incrementally for any new target user.

Incremental Unsupervised Domain-Adversarial Training of Neural Networks.

In the context of supervised statistical learning, it is typically assumed that the training set comes from the same distribution that draws the test samples. When this is not the case, the behavior of the learned model is unpredictable and becomes dependent upon the degree of similarity between the distribution of the training set and the distribution of the test set. One of the research topics that investigates this scenario is referred to as domain adaptation (DA). Deep neural networks brought dramatic advances in pattern recognition and that is why there have been many attempts to provide good DA algorithms for these models. Herein we take a different avenue and approach the problem from an incremental point of view, where the model is adapted to the new domain iteratively. We make use of an existing unsupervised domain-adaptation algorithm to identify the target samples on which there is greater confidence about their true label. The output of the model is analyzed in different ways to determine the candidate samples. The selected samples are then added to the source training set by self-labeling, and the process is repeated until all target samples are labeled. This approach implements a form of adversarial training in which, by moving the self-labeled samples from the target to the source set, the DA algorithm is forced to look for new features after each iteration. Our results report a clear improvement with respect to the non-incremental case in several data sets, also outperforming other state-of-the-art DA algorithms.

Unsupervised Domain Adaptation in the Wild via Disentangling Representation Learning

Most recently proposed unsupervised domain adaptation algorithms attempt to learn domain invariant features by confusing a domain classifier through adversarial training. In this paper, we argue that this may not be an optimal solution in the real-world setting (a.k.a. in the wild) as the difference in terms of label information between domains has been largely ignored. As labeled instances are not available in the target domain in unsupervised domain adaptation tasks, it is difficult to explicitly capture the label difference between domains. To address this issue, we propose to learn a disentangled latent representation based on implicit autoencoders. In particular, a latent representation is disentangled into a global code and a local code. The global code is capturing category information via an encoder with a prior, and the local code is transferable across domains, which captures the “style” related information via an implicit decoder. Experimental results on digit recognition, object recognition and semantic segmentation demonstrate the effectiveness of our proposed method.

A Systematic Study of Unsupervised Domain Adaptation for Robust Human-Activity Recognition

Wearable sensors are increasingly becoming the primary interface for monitoring human activities. However, in order to scale human activity recognition (HAR) using wearable sensors to million of users and devices, it is imperative that HAR computational models are robust against real-world heterogeneity in inertial sensor data. In this paper, we study the problem of wearing diversity which pertains to the placement of the wearable sensor on the human body, and demonstrate that even state-of-the-art deep learning models are not robust against these factors. The core contribution of the paper lies in presenting a first-of-its-kind in-depth study of unsupervised domain adaptation (UDA) algorithms in the context of wearing diversity -- we develop and evaluate three adaptation techniques on four HAR datasets to evaluate their relative performance towards addressing the issue of wearing diversity. More importantly, we also do a careful analysis to learn the downsides of each UDA algorithm and uncover several implicit data-related assumptions without which these algorithms suffer a major degradation in accuracy. Taken together, our experimental findings caution against using UDA as a silver bullet for adapting HAR models to new domains, and serve as practical guidelines for HAR practitioners as well as pave the way for future research on domain adaptation in HAR.

Does BERT need domain adaptation for clinical negation detection?

Classifying whether concepts in an unstructured clinical text are negated is an important unsolved task. New domain adaptation and transfer learning methods can potentially address this issue. We examine neural unsupervised domain adaptation methods, introducing a novel combination of domain adaptation with transformer-based transfer learning methods to improve negation detection. We also want to better understand the interaction between the widely used bidirectional encoder representations from transformers (BERT) system and domain adaptation methods. We use 4 clinical text datasets that are annotated with negation status. We evaluate a neural unsupervised domain adaptation algorithm and BERT, a transformer-based model that is pretrained on massive general text datasets. We develop an extension to BERT that uses domain adversarial training, a neural domain adaptation method that adds an objective to the negation task, that the classifier should not be able to distinguish between instances from 2 different domains. The domain adaptation methods we describe show positive results, but, on average, the best performance is obtained by plain BERT (without the extension). We provide evidence that the gains from BERT are likely not additive with the gains from domain adaptation. Our results suggest that, at least for the task of clinical negation detection, BERT subsumes domain adaptation, implying that BERT is already learning very general representations of negation phenomena such that fine-tuning even on a specific corpus does not lead to much overfitting. Despite being trained on nonclinical text, the large training sets of models like BERT lead to large gains in performance for the clinical negation detection task.