Search

Tissue semantic segmentation from Hematoxylin and Eosin (H&E)-stained Whole Slide Images (WSIs) is a highly effective technique in medical image analysis that has significantly impacted clinical practice and scientific research. Accurate delineation of tissue boundaries allows for precise identification of diseases and the formulation of successful treatment strategies. Nevertheless, the process for developing such a system presents considerable obstacles because of the intrinsic diversity in tissue structure, the existence of interference, and the extensive array of colour distributions. Several studies have investigated various deep learning architectural frameworks for the purpose of tissue segmentation. However, deep learning algorithms require a significant volume of data in order to successfully learn complex patterns and characteristics. The lack of consistent large-scale datasets hinders the progress of these models. To this end, semi-supervised learning strategies can facilitate accurate segmentation of WSIs in the absence of large annotated datasets. Moreover, there is a pressing need for the advancement of these technologies. The objective of this study is to design a semi-supervised learning technique for tissue semantic segmentation from WSI. The work employs a hybrid strategy that combines pseudo-label generation and consistency regularization to efficiently utilize the partially labeled dataset. Additionally, a novel pseudo-label refinement module is proposed to reduce the frequency of noisy pseudo-labels. Furthermore, there is a requirement for the development of a proficient feature extractor that can efficiently extract rich features from the images. This study tackles this problem by introducing a deep-learning segmentation framework, TS-Net, which can accurately segment tissue from histopathological images even with small datasets. The proposed model uses convolution and transformer layers to efficiently extract supplementary information from the image. The performance of the proposed model is comprehensively assessed both qualitatively and quantitatively using a publicly available dataset. The evaluation showcases competitive outcomes when compared to the other techniques currently available.

Electroencephalogram (EEG) signals are electrical signals generated by the activity between neurons in the brain and are already extensively applied for seizure prediction. Semi-supervised learning (SSL) has been applied in seizure prediction studies, as acquiring labeled EEG data can be time-consuming and costly. Current available SSL methods use far fewer labeled EEG samples than unlabeled EEG samples to train models. However, these methods have two limitations. First, the scarcity of EEG data limits the generalizability of the model. Second, the imbalance in the numbers of labeled and unlabeled EEG samples during model training leads to a potential distribution mismatch between the two sample types. Therefore, a novel semi-supervised seizure prediction method is designed in this study. First, an EEG data augmentation network is used to augment EEG signals at the signal level. Then, the labels of the original and augmented unlabeled EEG data are acquired from the spectrograms of EEG signals. Subsequently, the acquired labels are sharpened and the convex combination of labeled and unlabeled data is computed. Finally, deep pairwise representation alignment is performed to predict seizures. Several experiments are conducted using the proposed method on the Children’s Hospital Boston-Massachusetts Institute of Technology EEG database and the Kaggle Epilepsy Prediction Challenge EEG database. The results show that the proposed semi-supervised seizure prediction method is effective and reliable, given that it provides satisfactory results with as few as 25 labeled data per class when trained on these two datasets.

Occupancy information is essential for space management, energy efficiency, and in times of the COVID-19 pandemic, for crowd control. Obtaining labeled data is challenging due to hardware limitations, privacy considerations, and the required underlying costs. This study demonstrates the benefits of using Semi-Supervised Learning (SSL) for occupancy estimation in enclosed spaces; which requires less labeled data than other Machine Learning (ML) methods. It presents an empirical comparison between Supervised ML and SSL models in three real-life university classrooms (uncontrolled conditions). The data was collected for three weeks at each classroom using an in-house developed Internet of Things (IoT) device that measures air temperature, relative humidity, and atmospheric pressure. The ground truth records were gathered through manual logging of occupancy levels. Datasets’ sizes averaged 2350 entries with only 280 labeled instances per dataset. Support Vector Machine (SVM), Random Forest (RF), and Multi-Layer Perceptron (MLP) were used to define a performance baseline for supervised ML. Self-Training (ST) and Label Propagation (LP) were tested for SSL. ST achieved superior performance compared to baseline models (SVM, RF, MLP) with a highest average accuracy of 87.33% compared to SVM (86.66%). These results demonstrate the effectiveness of SSL for indirect occupancy estimation while reducing the need for extensive data collection and labeling.

Existing fully-supervised semantic segmentation methods have achieved good performance. However, they all rely on high-quality pixel-level labels. To minimize the annotation costs, weakly-supervised methods or semi-supervised methods are proposed. When such methods are applied to the infrared ship image segmentation, inaccurate object localization occurs, leading to poor segmentation results. In this paper, we propose an infrared ship segmentation (ISS) method based on weakly-supervised and semi-supervised learning, aiming to improve the performance of ISS by combining the advantages of two learning methods. It uses only image-level labels and a minimal number of pixel-level labels to segment different classes of infrared ships. Our proposed method includes three steps. First, we designed a dual-branch localization network based on ResNet50 to generate ship localization maps. Second, we trained a saliency network with minimal pixel-level labels and many localization maps to obtain ship saliency maps. Then, we optimized the saliency maps with conditional random fields and combined them with image-level labels to generate pixel-level pseudo-labels. Finally, we trained the segmentation network with these pixel-level pseudo-labels to obtain the final segmentation results. Experimental results on the infrared ship dataset collected on real sites indicate that the proposed method achieves 71.18% mean intersection over union, which is at most 56.72% and 8.75% higher than the state-of-the-art weakly-supervised and semi-supervised methods, respectively.

In the real world, there are only a small amount of data with labels. To make full use of the potential structural information of unlabeled data to train a better classifier, researchers have proposed many semi-supervised learning algorithms. Among these algorithms, self-training is one of the most widely used semi-supervised learning frameworks due to its simplicity. How to select high-confidence samples is a crucial step for self-training. If the misclassified samples are selected as high-confidence samples, this error will be amplified in the iterative process, which affects the performance of the final classifier. To alleviate the impact of this problem, this paper proposes a self-training algorithm with block-similar neighbor editing (STBSNE). STBSNE calculates the distance between samples by the block-based dissimilarity measure, which improves the classification performance on high-dimensional data sets. STBSNE defines the block-estimated neighbor relationship, builds the block-estimated neighbor relationship graph, and proposes the block estimated neighbor editing method to identify outliers and noise points, and edits them to improve the quality of the high-confidence sample selected. Experimental results on 16 benchmark data sets verify the superior performance of the proposed STBSNE compared with seven state-of-the-art algorithms.

We present a novel data-efficient semi-supervised framework to improve the generalization of image captioning models. Constructing a large-scale labeled image captioning dataset is expensive in terms of labor, time, and cost. In contrast to manually annotating all the training samples, separately collecting uni-modal datasets is immensely easier, e.g., a large-scale image dataset and a sentence dataset. We leverage such massive unpaired image and caption data upon standard paired data by learning to associate them. To this end, our novel semi-supervised learning method assigns pseudo-labels to unpaired images and captions in an adversarial learning fashion, where the joint distribution of image and caption is learned. This approach shows noticeable performance improvement even in challenging scenarios, including out-of-task data and web-crawled data. We also show that our proposed method is theoretically well-motivated and has a favorable global optimal property. Our extensive and comprehensive empirical results on captioning datasets, followed by a comprehensive analysis of the scarcely-paired COCO dataset, demonstrate the consistent effectiveness of our method compared to competing ones.

Segmentation of internal structures in echocardiography enables quantitative measurements of the cardiac size, shape, and function, thereby facilitating automated diagnosis and treatment planning. Semi-supervised learning alleviates the demand for large amounts of annotations in deep learning. The existing interpolation-based data augmentation (perturbation) methods demonstrate significant potential, yet they solely focus on operating pixels in the spatial domain, overlooking considerations in the frequency domain. Our proposed interpolation method at the frequency domain level operates on a level where the spatial arrangement of pixels is less relevant. It is capable of generating potent and diverse perturbations while concurrently circumventing disruption to the spatial structure of the image, thereby augmenting the robustness of semi-supervised frameworks. Specifically, we transform annotated and unannotated images into the frequency domain separately. Then, half of the elements in each frequency domain map are randomly selected and mixed with each other. Finally, the resulting mixed frequency domain images are transformed back into the spatial domain to generate mixed images. Furthermore, we refine the spatial domain CutMix method by devising a cutting strategy that adheres to jigsaw partitioning rules, facilitating the model’s perception of local cardiac anatomical structures. Extensive experiments on two public echocardiography segmentation datasets demonstrate the effectiveness of the proposed method.

Semisupervised change detection (CD) methods have garnered increasing attention due to their capacity to alleviate the dependency of fully-supervised methods on a large number of pixel-level labels. These methods predominantly leverage generative adversarial network architecture and consistency regularization technology. However, they encounter challenges associated with background noise from cross-temporal images. In this article, we propose a novel multilevel consistency-regularization-based semisupervised CD approach that incorporates Fourier-based frequency transformation and a reliable pseudolabel selection scheme. Specifically, we replace the low-frequency spectrum of one temporal image with a frequency domain transformation derived from the corresponding image in the same bitemporal remote sensing image pair, enhancing the model's capability to discern meaningful changes amidst background noise, thereby contributing to more robust CD. Furthermore, excessively high pseudolabel thresholds in consistency regularization methods may result in the underutilization of valuable unlabeled data. To address this issue, we design a straightforward sigmoid-like function to dynamically adjust the selection threshold for the reliable pseudolabel selection scheme. This strategy takes into consideration the learning status throughout the entire training process, ensuring more effective utilization of unlabeled information. We demonstrate significant performance improvements across three widely-used public datasets, namely, LEVIR-CD, WHU-CD, and CDD. Notably, on the three datasets with only 1% labeled data, our method achieved an $\text{IoU}^{c}$ of 71.29%, 63.90%, and 51.00%, outperforming existing state-of-the-art methods by 2.84%, 1.21%, and 0.98%, respectively. These results robustly substantiate the effectiveness of our approach, showcasing its potential in scenarios where labeled data is limited.

Goal: Deep learning techniques have made significant progress in medical image analysis. However, obtaining ground truth labels for unlabeled medical images is challenging as they often outnumber labeled images. Thus, training a high-performance model with limited labeled data has become a crucial challenge. Methods: This study introduces an underlying knowledge-based semi-supervised framework called UKSSL, consisting of two components: MedCLR extracts feature representations from the unlabeled dataset; UKMLP utilizes the representation and fine-tunes it with the limited labeled dataset to classify the medical images. Results: UKSSL evaluates on the LC25000 and BCCD datasets, using only 50% labeled data. It gets precision, recall, F1-score, and accuracy of 98.9% on LC25000 and 94.3%, 94.5%, 94.3%, and 94.1% on BCCD, respectively. These results outperform other supervised-learning methods using 100% labeled data. Conclusions: The UKSSL can efficiently extract underlying knowledge from the unlabeled dataset and perform better using limited labeled medical images.

This paper presents GReAT (Graph Regularized Adversarial Training), a novel regularization method designed to enhance the robust classification performance of deep learning models. Adversarial examples, characterized by subtle perturbations that can mislead models, pose a significant challenge in machine learning. Although adversarial training is effective in defending against such attacks, it often overlooks the underlying data structure. In response, GReAT integrates graph-based regularization into the adversarial training process, leveraging the data’s inherent structure to enhance model robustness. By incorporating graph information during training, GReAT defends against adversarial attacks and improves generalization to unseen data. Extensive evaluations on benchmark datasets demonstrate that GReAT outperforms state-of-the-art methods in robustness, achieving notable improvements in classification accuracy. Specifically, compared to the second-best methods, GReAT achieves a performance increase of approximately 4.87% for CIFAR-10 against FGSM attack and 10.57% for SVHN against FGSM attack. Additionally, for CIFAR-10, GReAT demonstrates a performance increase of approximately 11.05% against PGD attack, and for SVHN, a 5.54% increase against PGD attack. This paper provides detailed insights into the proposed methodology, including numerical results and comparisons with existing approaches, highlighting the significant impact of GReAT in advancing the performance of deep learning models.

Deep neural networks (DNNs) have proven highly effective in various computational tasks, but their success depends largely on access to large datasets with accurate labels. Obtaining such data may be challenging and costly in real-world scenarios. Common alternatives, such as the use of search engines and crowdsourcing, often result in datasets with inaccurately labeled, or “noisy,” data. This noise may significantly reduce the ability of DNNs to generalize and maintain reliability. Traditional methods for learning with noisy labels mitigate this drawback by training DNNs selectively on reliable data, but they often underutilize available data. Although data augmentation techniques are useful, they do not directly solve the noisy label problem and are limited in such contexts. This paper proposes a confidence-guided Mixup named ConfidentMix, which is a data augmentation strategy based on label confidence. Our method dynamically adjusts the intensity of data augmentation according to label confidence, to protect DNNs from the detrimental effects of noisy labels and maximize the learning potential from the most reliable portions of the dataset. ConfidentMix represents a unique blend of label confidence assessment and customized data augmentation, and improves model resilience and generalizability. Our results on standard benchmarks with synthetic noise, such as CIFAR-10 and CIFAR-100, demonstrate the superiority of ConfidentMix in high-noise environments. Furthermore, extensive experiments on Clothing1M and mini-WebVision have confirmed that ConfidentMix surpasses state-of-the-art methods in handling real-world noise.

The Internet provides a platform for sharing services, and web service brokers help users to choose the suitable service among similar services based on ranking. The quality of service is important in evaluating the services the user needs. However, finding a quality-based data label in many fields can be time-consuming and difficult. Thus, machine learning is required to classify and choose the best service in this field. The selection process is done through analysis and recommendations by the system. This article introduces the SSL-WSC algorithm, which classifies unlabeled data through semi-supervised self-training learning using a small amount of labeled data. This algorithm labels the data using a two-step method of calculating a score for each service and dynamic thresholding. The quality features of web services obtained from the QWS dataset were used to evaluate the performance of the proposed algorithm. The experimental results in different scenarios showed that using proposed semi-supervised learning algorithms to create classification models led to better results, so it improved the F1-score, accuracy, and precision, on average, by 11.26%, 9.43% and 9.53%, respectively, as compared to the supervised method.

Electromagnetic field exposure (EMF) has grown to be a critical concern as a consequence of the ongoing installation of fifth-generation cellular networks (5G). The lack of measurements makes it difficult to accurately assess the EMF in a specific urban area, as Spectrum cartography (SC) relies on a set of measurements recorded by spatially distributed sensors for the generation of exposure maps. However, when the spatial sampling rate is limited, significant estimation errors occur. To overcome this issue, the exposure map estimation is addressed as a missing data imputation task. We compute a convolutional neural tangent kernel (CNTK) for an infinitely wide convolutional neural network whose training dynamics can be completely described by a closed-form formula. This CNTK is employed to impute the target matrix and estimate EMF exposure from few sensors sparsely located in an urban environment. Experimental results show that the kernel, even when only sparse sensor data are available, can produce accurate estimates. It is a promising solution for exposure map reconstruction that does not require large training sets. The proposed method is compared with other deep learning approaches and Gaussian Process regression.

Breast tumor is a common female physiological disease, and the malignant tumor is one of the main fatal diseases of women. Accurate examination and assessment of tumor shape can facilitate subsequent treatment and improve the cure rate. With the development of deep learning, automatic detection systems are designed to assist doctors in diagnosis. However, the blurry edges, poor visual quality, and irregular shapes of breast tumors pose significant challenges to design a highly efficient detection system. In addition, the lack of publicly available labeled data is a major obstacle in developing highly accurate and robust deep learning models for breast tumor detection. To overcome the aforementioned issues, we propose SRU-PMT+, a pseudo-label reusing Mean-Teacher architecture based on squeeze-and-excitation residual (SE-Res) attention. We utilize the proposed segmentation network, SRU-Net++, to generate pseudo-labels for unlabeled data, and guide the learning of the student model using the generated pseudo-labels and groundtruth, improving the accuracy and robustness of the model. Our proposed semi-supervised method has been rigorously evaluated on the available labeled dataset, i.e., Breast Ultrasound Images (BUSI) dataset. Results show that our proposed method outperforms current segmentation methods and has good performance. Importantly, our strategy of reusing pseudo-labels improves the performance of breast tumor segmentation.

The Vision Transformer (ViT) model serves as a powerful model to capture and comprehend global information, particularly when trained on extensive datasets. Conversely, the Convolutional Neural Network (CNN) model is beneficial to training with small datasets for retaining essential local information. Inspired by these properties of ViT and CNN models, we introduce a hybrid framework that smoothly increases the cross-domain generalization in Semi-supervised Domain Adaptation (SSDA). To achieve this goal, we first train the ViT model on abundant labeled source data, while the CNN model is trained on a few labeled target samples. Then, these models dynamically exchange their knowledge for potential generalization to unlabeled target data via the proposed method, named Dual Dynamic Consistency Regularization (D2CR). Specifically, the ViT model provides its pseudo labels to update the global perspective for the CNN model. Similarly, the CNN model offers pseudo labels to complement the local perspective for the ViT model. The previous methods use a fixed threshold algorithm for the pseudo-labeling process. However, we utilize the dynamic threshold strategy to create pseudo labels for the bi-directional consistency regularization learning between the ViT and CNN models. We verify our approach across several SSDA benchmark datasets. The outstanding experimental results provide strong evidence of the effectiveness and superiority of our approach over previous state-of-the-art SSDA methods.

We present Polite Teacher, a simple yet effective method for the task of semi-supervised instance segmentation. The proposed architecture relies on the Teacher-Student mutual learning framework. To filter out noisy pseudo-labels, we use confidence thresholding for bounding boxes and mask scoring for masks. The approach has been tested with CenterMask, a single-stage anchor-free detector. Tested on the COCO 2017 val dataset, our architecture significantly (approx. +8 pp. in mask AP) outperforms the baseline at different supervision regimes. To the best of our knowledge, this is one of the first works tackling the problem of semi-supervised instance segmentation and the first one devoted to an anchor-free detector. The code is available: github.com/AI-Clearing/PoliteTeacher.

With the rapid development of information technology, efficient multi label classification of massive data is one of the important tasks of big data systems. Semi supervised learning algorithm is an effective data classification method, currently mainly applied to the classification of single label data. This article proposes a multi label dynamic semi supervised learning algorithm based on dual selection criteria. The algorithm mainly establishes dual selection criteria for multi label pseudo labeled samples based on the COIN structure and K-nearest neighbor algorithm. A novel pseudo labeled sample selection method is designed, which improves the robustness and accuracy of the algorithm and effectively solves the problem of not considering sample correlation when selecting pseudo labeled samples. On this basis, by adding a performance evaluation mechanism to the model, the model can dynamically and adaptively extract pseudo labeled samples, improving the training speed and accuracy of the model. This article selected four convincing public test datasets for experiments, and the experimental results showed that the proposed semi supervised learning method has improved in multiple indicators such as robustness, accuracy, and training efficiency compared to current mainstream methods.

Deep learning has attracted wide attention recently because of its excellent feature representation ability and end-to-end automatic learning method. Especially in clinical medical imaging diagnosis, the semi-supervised deep learning model is favored and widely used because it can make maximum use of a limited number of labeled data and combine it with a large number of unlabeled data to extract more information and knowledge from it. However, the scarcity of medical image data, the vast image size, and the instability of image quality directly affect the model’s robustness, generalization, and image classification performance. Therefore, this paper proposes a new semi-supervised learning model, which uses quadratic neurons instead of traditional neurons, aiming to use quadratic convolution instead of the conventional convolution layer to improve the feature extraction ability of the model. In addition, we introduce two Dropout layers and two fully connected layers at the end of the model to enhance the nonlinear fitting ability of the network. Experiments on two large medical public data sets - ISIC 2019 and Retinopathy OCT - show that our method can improve the model’s generalization performance and image classification accuracy.

This study proposes a novel semi-supervised multi-label emotion classification approach for French tweets based on pseudo-labeling. Human subjectivity in emotional expression makes it difficult for a machine to learn. Therefore, it necessitates training supervised learning models on large datasets annotated by multiple annotators. However, creating such datasets can be costly and time-consuming. Moreover, aggregating annotations from multiple annotators to capture their collective emotional state adds complexity to the task. Semi-supervised learning techniques have shown effectiveness with limited datasets. Furthermore, Large language Models (LLMs), particularly Chat-GPT, have demonstrated superior annotation accuracy compared to annotations obtained from crowdsourcing platforms, when both are evaluated against expert-annotated data. This work introduces a novel approach for multi-label emotion classification of French tweets by leveraging pseudo-labels generated through Chat-GPT, a robust large language model. Using zero-shot, one-shot, and few-shot learning techniques, Chat-GPT annotates the unlabelled part of our dataset. These Chat-GPT-annotated pseudo-labels are then merged with manual annotations, facilitating the training of a multi-label emotion classification model via semi-supervised learning. Furthermore, within the context of our research, we present a new French tweet dataset, containing testimonials from people affected by an urban industrial incident. This dataset features 2,350 tweets, each manually annotated by three human annotators based on 8 pre-identified emotions. Benchmark results are presented for multi-label emotion classification models employing both fully supervised and semi-supervised approaches with pseudo-labeling. Our findings demonstrate the superiority of our approach for multi-label emotion classification over standard pseudo-labeling and an established baseline.

In response to the growing demand for 3D object detection in applications such as autonomous driving, robotics, and augmented reality, this work focuses on the evaluation of semi-supervised learning approaches for point cloud data. The point cloud representation provides reliable and consistent observations regardless of lighting conditions, thanks to advances in LiDAR sensors. Data annotation is of paramount importance in the context of LiDAR applications, and automating 3D data annotation with semi-supervised methods is a pivotal challenge that promises to reduce the associated workload and facilitate the emergence of cost-effective LiDAR solutions. Nevertheless, the task of semi-supervised learning in the context of unordered point cloud data remains formidable due to the inherent sparsity and incomplete shapes that hinder the generation of accurate pseudo-labels. In this study, we consider these challenges by posing the question: “To what extent does unlabelled data contribute to the enhancement of model performance?” We show that improvements from previous semi-supervised methods may not be as profound as previously thought. Our results suggest that simple grid search hyperparameter tuning applied to a supervised model can lead to state-of-the-art performance on the ONCE dataset, while the contribution of unlabelled data appears to be comparatively less exceptional.

Defect patterns exhibited in wafer bin maps (WBMs) can provide essential clues about critical process failures to field engineers. In modern manufacturing processes, the automatic WBM defect pattern classification is critical for yield improvement. Although it is difficult to collect sufficient labels while a lot of unlabeled data is given, most existing studies have mainly used only labeled WBM data. Moreover, the unlabeled out-of-distribution (OOD) WBMs are inevitably collected. It degrades the performance of semi-supervised models. To this end, we propose a method for safe wafer bin map classification with self-supervised contrastive learning (SWaCo) to effectively exploit the unlabeled data with OOD. We propose a loss function to utilize label information in the pre-training step to learn more suitable representations for the downstream WBM classification task. The negative labeled examples of the same class as the anchor are used for additional positive examples. Moreover, we investigate proper data augmentation for WBM self-supervised contrastive learning. To evaluate the performance and the applicability of the proposed method, experiments are conducted on a public benchmark WBM dataset, WM-811K. The results demonstrate that the proposed method achieves better classification accuracy than existing methods, especially when only a small amount of class information is given. [ABSTRACT FROM AUTHOR]

Automatic defect classification (ADC) systems automatically classify defects that inevitably occur during semiconductor manufacturing processes. ADC is the beginning of defect management that increases the yield of semiconductor chip production, and prevents accidents in the process. It takes a lot of engineer’s labor to classify defects, but ADC can be the answer to classify all defects at low cost. ADC employs the defect image of a wafer surface, captured using scanning electron microscopy (SEM). SEM images can feature a variety of backgrounds based on the defect position on the wafer and the process steps. The manual classification of SEM images involves significant labor costs related to hiring experienced engineers. Despite recent ADC studies reporting good performance, the lack of labeled images and various backgrounds make it difficult to apply ADC in actual manufacturing processes. To address this issue, automated defect classification with defect localization is proposed herein. To this end, a classification model is specifically designed for reducing the effect of varying backgrounds using defect localization. Defect localization uses an object detection model to provide the region information of defects in SEM images. We aimed to design a classification model and defect detection model using semi-supervised learning to reduce labeling costs. Experimental results indicate that the classification performance, over 15 classes, is improved by 12.56% (9.82%p), as compared with that of supervised models. [ABSTRACT FROM AUTHOR]

Traditional inverse Q filtering methods for post-stack seismic data attenuation compensation (AC) have the drawback of instability or under-compensation. Besides, the quality factor (Q) should be known as a prerequisite, which is commonly estimated using the attribute difference between the reference and observed wavelets of pre-stack vertical seismic profile (VSP) data. The alternating iterative AC and Q estimation method is also researched for post-stack data, while the instability or huge computation becomes a defect. In this letter, we propose a simultaneous AC and Q estimation method for nonstationary post-stack seismic data based on semi-supervised learning. Specifically, we choose the long short-term memory algorithm which is sensitive to time series and can characterize seismic signal nonlinearly with high accuracy. The proposed AC and Q estimation method employs the Q information from well-logs and the compensated high-resolution data for supervised learning, and uses nonstationary seismic data beyond wells for self-supervised learning, without the wavelet extraction procedure. The synthetic data analysis and field data applications prove the feasibility of the designed semi-supervised method in improving the vertical resolution and Q estimation. The field data impedance inversion after AC further demonstrates its effectiveness. [ABSTRACT FROM AUTHOR]

Deep learning technologies have been successfully applied to hyperspectral (HS) image classification with remarkable performance. However, compared with traditional machine learning methods, neural networks usually need more data. In remote sensing (RS) research, obtaining a large number of labeled HS data is very difficult and expensive work. Simultaneously, the distribution of feature information is bound to be unbalanced, and tends to conform to the long tail. At present, the neighborhood information of unlabeled samples is usually ignored in HS image classification tasks based on semi-supervised learning. In this letter, we propose a new semi-supervised long-tail learning framework based on spatial neighborhood information (SLN-SNI), which can complete the HS image classification task under unbalanced small sample data. Specifically, a new semi-supervised learning strategy is proposed. On this basis, a new method to determine the label of unlabeled samples based on spatial neighborhood information (SNI) is proposed. The coarse classification results divided into three situations are judged again, and the accuracy of pseudo labels is improved. The performance of the proposed method is tested on three public HS image datasets. Compared with the current advanced methods have achieved a certain improvement. [ABSTRACT FROM AUTHOR]

The semi-supervised face recognition problem has become a popular research topic in recent years. However, one common and important situation, in which the unlabeled data is shallow, has rarely been considered in most existing works. In this paper, shallow data means there are only few images per identity. In the unlabeled shallow situation, the existing semi-supervised face recognition methods generally do not work well. Thus, how to effectively utilize the unlabeled shallow face data for improving face recognition performance is an important issue. In this paper, we propose a novel semi-supervised face recognition method, namely VirFace $^{\infty} $ , to enhance the face recognition performance effectively with the unlabeled shallow data. VirFace $^{\infty} $ consists of VirClass and VirDistribution components. In VirClass, we inject the unlabeled data as virtual classes into the feature space to enlarge the inter-class distance. In VirDistribution, we predict the distribution of each virtual class, namely virtual distribution, and then enhance the inter-class discriminativeness by enlarging the distances between the labeled features and the virtual distributions. To the best of our knowledge, we are among the first to tackle the face recognition problem on unlabeled shallow face data. Extensive experiments demonstrate the superiority of our proposed method. [ABSTRACT FROM AUTHOR]

Semi-Supervised Learning (SSL) is an influential derivative that allows humans to uncover invisible knowledge, potentially substituting it for extensive labeling data. Despite the optimism generated by the availability of unlabeled data, its potential unreliability can result in numerous unknown security risks. Assailants may covertly contaminate data, leading to potentially catastrophic and unpredictable outcomes. We investigate poisoning attacks in triangular manifolds to understand how SSL models defend against attacks resulting from small perturbations. By inserting tiny amounts of artificially modified samples totaling 2% of the entire training set, we can deceive classification models into altering the prediction results of arbitrary categories. In addition, considering that the poisoned data in practical scenarios belong to a minority sample attack, which is typically only about 0.1%-2% of the total data, we employed outlier detection to examine all inserted instances and discovered that it could bypass discovery. Our poisoning strategy can work across multiple datasets, models, and application domains of images and network traffic. Experimental results prove that our proposed method is effective on at least seven semi-supervised models. The declining ratio of model detection accuracy of autoencoder with confidence (ConAE) is 52.55% at the lowest cost. Another persuasive result is that our model poisoning exceeds the state-of-the-art methods in the image domain. The extended conclusion corroborates that the more accurate classification models do not have a corresponding improvement in their ability to resist interference, which also provides a new standard for testing model robustness. [ABSTRACT FROM AUTHOR]

Conversational emotion recognition is a crucial research topic in human-computer interactions. Due to the heavy annotation cost and inevitable label ambiguity, collecting large amounts of labeled data is challenging and expensive, which restricts the performance of current fully-supervised methods in this domain. To address this problem, researchers attempt to distill knowledge from unlabeled data via semi-supervised learning. However, most of these semi-supervised methods ignore multimodal interactive information, although recent works have proven that such interactive information is essential for emotion recognition. To this end, we propose a novel framework to seamlessly integrate semi-supervised learning with multimodal interactions, called “Semi-supervised Multi-modal Interaction Network (SMIN)”. SMIN contains two well-designed semi-supervised modules, “Intra-modal Interactive Module (IIM)” and “Cross-modal Interactive Module (CIM)” to learn intra- and cross-modal interactions. These two modules leverage additional unlabeled data to extract emotion-salient representations. To capture additional contextual information, we utilize the hierarchical recurrent networks followed with the hybrid fusion strategy to integrate multimodal features. These multimodal features are further utilized for conversational emotion recognition. Experimental results on four benchmark datasets (i.e., IEMOCAP, MELD, CMU-MOSI and CMU-MOSEI) demonstrate that SMIN succeeds over existing state-of-the-art strategies on emotion recognition. [ABSTRACT FROM AUTHOR]

In this paper, we aim to improve the performance of in-the-wild Facial Expression Recognition (FER) by exploiting semi-supervised learning. Large-scale labeled data and deep learning methods have greatly improved the performance of image recognition. However, the performance of FER is still not ideal due to the lack of training data and incorrect annotations (e.g., label noises). Among existing in-the-wild FER datasets, reliable ones contain insufficient data to train robust deep models while large-scale ones are annotated in lower quality. To address this problem, we propose a semi-supervised learning algorithm named Progressive Teacher (PT) to utilize reliable FER datasets as well as large-scale unlabeled expression images for effective training. On the one hand, PT introduces semi-supervised learning method to relieve the shortage of data in FER. On the other hand, it selects useful labeled training samples automatically and progressively to alleviate label noise. PT uses selected clean labeled data for computing the supervised classification loss and unlabeled data for unsupervised consistency loss. Experiments on widely-used databases RAF-DB and FERPlus validate the effectiveness of our method, which achieves state-of-the-art performance with accuracy of 89.57% on RAF-DB. Additionally, when the synthetic noise rate reaches even 30%, the performance of our PT algorithm only degrades by 4.37%. [ABSTRACT FROM AUTHOR]

Responding to the challenge of efficiently leveraging Open-source Threat Intelligence Feeds (OTIFs) to enhance organizational security, this paper presents an innovative approach to automated threat identification using machine learning. Current machine learning models require large volumes of high-quality annotated data, which is a major obstacle given the inherent scarcity of such data in OTIFs. Bridging this research gap, we propose a novel semi-supervised learning strategy that capitalizes on both labeled and unlabeled data to automate threat identification. Our unique contribution is an advanced iteration of the GAN-BERT framework, which incorporates a self-supervised contrastive objective function to fine-tune the BERT language model. Our experiments show that this method outperforms the original BERT and GAN-BERT, achieving a 3-12% F1-score improvement on various OTIFs datasets. Furthermore, we present an efficient method for selecting hard negatives during training, which further enhances our model’s performance. This innovative approach significantly advances the field of automated threat detection, reducing reliance on human supervision and effectively addressing the issue of limited annotated data. Thus, it offers a robust solution for strengthening security posture through proactive decision making in Cyber Threat Intelligence.

Non-intrusive load monitoring is a novel and cost-effective technology for monitoring details of electricity consumption and identifying the operating status of appliances. It supports the construction of the energy internet and big data on electricity consumption in smart cities. However, one of the most challenging problems in this area is that machine learning algorithms often require large amounts of labeled data. In this paper, a non-intrusive load monitoring model based on the Self-supervised Regularization is proposed. The model reduces the pre-processing stage compared to the traditional methods. We make full use of the unlabeled data by using them to generate proxy labels to participate in the model training together with the true labels. We performed experiments on the common data set PLAID to compare performance with the existing method Mean Teacher and CoMatch. The experimental results show that: 1) when using all labeled data, the model with self-supervised regularization significantly improves the traditional supervised classifier with a recognition accuracy of 0.965; 2) when coupled with unlabeled data, our model produces good semi-supervised performance. It is highly competitive with current state-of-the-art Mean Teacher and Contrastive Learning.

The rapid spread of biased information and disinformation has intensified in recent times due to the increased reliance on the Internet and social media platforms as primary sources of information. This issue is of particular concern in the fields of medicine and healthcare, given the critical nature of decisions and understandings in these areas. While medical experts can mitigate the repercussions of misinformation by evaluating questionable websites, this approach is time-consuming. Consequently, there is a pressing need to develop software solutions that can automate the detection of misleading information. This paper presents CO-training and Active Learning-based framework for Finding Low-credibility web Addresses in the MEdical field (COAL4FLAME), a novel system designed to analyze health-related websites and identify misinformation. The system integrates results from multiple estimators to reach a comprehensive conclusion. COAL4FLAME uses semi-supervised learning strategies, such as multi-view learning, co-training, and active learning, to address the challenge of limited labeled data, a major strength of the presented proposal. Medical experts have rigorously tested and evaluated the system using a selected set of websites.

Classification of brain tumors from the Magnetic Resonance Imaging (MRI) is a vital and challenging task for brain tumor diagnosis. Despite favorable results, from current Deep Learning (DL) methods used for the classification of brain tumors, the approaches are also used to cope with inconsistencies where the visual representation of non-brain tumor and brain tumor regions are equivalent. In this article, a novel semi-supervised DL-based approach is proposed for the classification of three types of brain tumors as well as no tumors. The proposed semi-supervised approach, Semi-Supervised Brain Tumor Classification Network (SSBTCNet), profits from the combination of an unsupervised AutoEncoder (AE) with supervised classification networks, in which an AE for learning concealed descriptors and a multi-layer perceptron-based classifier are trained concurrently. When compared to training the AE and classifier separately, the proposed semi-supervised learning fundamentally aids in tuning the learning of concealed descriptors for the objective of classification, resulting in improved brain MRI classification performance. To make SSBTCNet efficient, enhanced instances are generated by using a fuzzy-logic-based method to train and test the system. Also, a second approach, for the classification of the brain tumor as well as no tumor, with augmented unlabelled data is trained and tested which applies five types of augmentations to the images of the dataset to make the system robust from the change in different scales, flips and orientation. The output of the proposed method is tested on various datasets (with and without data augmentation as well as with and without labeled data) where the accuracy rate specifies that the proposed approach outperforms the other brain tumor classification approaches used in the literature.

Deep neural networks (DNNs) are highly vulnerable to neural Trojan attacks. To carry out such an attack, an adversary retrains a DNN with poisoned data or modifies its parameters to produce incorrect output. These attacks can remain unnoticed until triggered by a specific pattern in the input, making detection challenging. In this article, we propose a novel semi-supervised adversarially learned meta-classifier (SESALME) to detect if a target model has been trojaned. Unlike previous Trojan detection methods, SESALME assumes that the defender has no knowledge of the attack mechanisms, and no access to training data, poisoned data, or parameters/layers of a target model. In the absence of poisoned data and knowledge of the attack mechanisms, we use a set of shadow models to emulate normal behavior of the target model. Having learned the normal behavior of the target model, SESALME then uses one-class learning, implemented within a semi-supervised generative adversarial network (GAN), to detect abnormal behavior of a model to be investigated, if any. Behavior that deviates from the learned normal behavior indicates a high likelihood that the model is trojaned. We compare the performance of SESALME with that of state-of-the-art neural Trojan detectors using popular datasets such as MNIST, CIFAR-10, and SC. Experimental results show that SESALME outperforms state-of-the-art Trojan detection methods in terms of detection performance and inference time in almost all cases, while being attack-agnostic and requiring no access to training data, poisoned data, or parameters of the target model.

Hyperspectral data classification is one of the fundamental problems in remote sensing. Several algorithms based on supervised machine learning have been proposed to address it. The performance, however, of the proposed algorithms is inherently dependent on the amount and quality of annotated data. Due to recent advances in hyperspectral imaging and autonomous (unmanned) aerial vehicles collecting new hyperspectral data is an easy task. Annotating those data, however, is a tedious, time-consuming and costly task requiring the in-situ presence of human experts. One way to loosen the requirement of a large number of annotated data is the shift to semi-supervised learning combined with highly sample-efficient tensor-based neural networks. This study provides a comprehensive experimental analysis of the performance of a variety of graph-based semi-supervised learning techniques combined with tensor-based neural network embeddings for the problem of hyperspectral data classification. Experimental results suggest that the combination of tensor-based neural network embeddings with graph-based semi-supervised learning can significantly improve the classification results minimizing human annotation effort.

As a critical intermediate material in the iron and steel industry, strip steel will inevitably have various surface defects during its processing, which directly affects the service performance and life of the material. Therefore, the classification technology of strip surface defects has always been the focus of research. Currently, combining computer vision with deep learning is often used to classify the surface defects of strip steel, which usually runs in full supervision mode. However, the performance of the complete supervision method depends mainly on the quality and quantity of labeled samples. At the same time, in industrial scenes, there are few labeled samples available, and most even have no labels, which seriously restricts the performance of the traditional full supervision model. This paper introduces the idea of semi-supervised learning, and a new semi-supervised classification model of strip surface defects is proposed to alleviate the degradation of model classification performance caused by insufficient labeled samples. Specifically, a new image synthesis model (ISM) is proposed in this paper. By improving the loss function of the discriminator, the generated false samples are more realistic. In addition, this paper also presents a double uncertainty weighting technique (DUW), which weighs the loss of misclassified samples in a more detailed way, thus realizing fine adjustment of the model. This method can fully mine the potential feature information in unlabeled samples and further improve the performance and generalization ability of the model. In this paper, we use the NEU-CLS dataset to test our model. When only 10% and 90% of labeled and unlabeled samples are used for training, the classification accuracy reaches 91.14%, fully proving this method’s practicability and superiority.

Manga (Japanese comics) are commonly drawn with black ink on paper. Colorization of manga pages can enrich the visual content and provide a better reading experience. However, the existing colorization approaches are not sufficiently robust. In this paper, we propose a two-stage approach for manga page colorization that supports sampling and color modification with color hints. In the first step, we employ the Pixel2Style2Pixel architecture to map the black-and-white manga image into the latent space of StyleGAN pretrained on the highly blurred colored manga images that we call Coloring Latent Space. The latent vector is automatically or manually modified and fed into the StyleGAN synthesis network to generate a coloring draft that sets the overall color distribution for the image. In the second step, heavy Pix2Pix-like conditional GAN fuses the information from the coloring draft and user-defined color hints and generates the final high-quality coloring. Our method partially overcomes the multimodality of the considered problem and generates diverse but consistent colorings without user input. The visual comparison, the quantitative evaluation with Frechet Inception Distance, and the qualitative evaluation via Mean Opinion Score exhibit the superiority of our approach over the existing state-of-the-art manga pages colorization method.

Time-series data with missing values are a common occurrence in various fields, including healthcare, meteorology, and robotics. The process of imputation aims to fill in the missing values with valid values. Most imputation methods implicitly train models due to the presence of missing values. In this paper, we propose Random Drop Imputation with Self-training (RDIS), a novel training method for time-series data imputation models. In RDIS, we generate extra missing values by applying a random drop to the observed values in incomplete data. We can explicitly train the imputation models by filling in the missing values. Moreover, we utilize self-training with pseudo values to exploit the original missing values. To enhance the quality of pseudo values, we set a threshold and filter them based on entropy calculation. To evaluate the effectiveness of RDIS for imputing time-series data, we test it across several imputation models and obtain competitive results on three real-world datasets.

Precise and robust image segmentation is one of the most important steps in supervised deep learning-applied studies. Especially in the medical field, image segmentation requires an enormous time and professionals with clinical knowledge. Although there are constant attempts for automatic and semi-automatic image segmentation algorithm development, acquiring not only clinically accurate but also precise pixel-level annotations for medical images remains insufficient. This article presents a semi-supervised learning method with a novel fact-forcing process, referred to as FFSS, to reduce the labeling cost while improving the prediction accuracy for medical image segmentation. FFSS includes two components: a pre-trained teacher and a student that would be trained, iteratively. In each iteration, the teacher first generates a pseudo-label for each image in an unlabeled set, the student is then trained on the pseudo-labeled set and sends feedback to update the teacher. A fact-forcing process is designed to improve the quality of the student model using a labeled set. We have comprehensively evaluated our method on both three-dimensional binary segmentation and two-dimensional multi-class segmentation. The evaluation results demonstrate significant accuracy improvements of FFSS compared with the state-of-the-art semi-supervised methods. Due to the fact-forcing process, the proposed method consistently outperforms the other ones under various labeled data ratios for all benchmark datasets, including left atrium MRI, pancreas CT, ACDC MRI, and OCT. By refining the quality of student feedback with complementary supervised training, the proposed FFSS shows robustness under labeled data scarcity for diverse types of medical images.

Metaverse is expected to rely on massive Internet of Things (IoT) connections so it inherits various security threats from the IoT network and also faces other sophisticated attacks related to virtual reality technology. As traditional security approaches show various limitations in the large-scale distributed metaverse, this paper proposes MetaCIDS, a novel collaborative intrusion detection (CID) framework that leverages metaverse devices to collaboratively protect the metaverse. In MetaCIDS, a federated learning (FL) scheme based on unsupervised autoencoder and an attention-based supervised classifier enables metaverse users to train a CID model using their local network data, while the blockchain network allows metaverse users to train a machine learning (ML) model to detect intrusion network flows over their monitored local network traffic, then submit verifiable intrusion alerts to the blockchain to earn metaverse tokens. Security analysis shows that MetaCIDS can efficiently detect zero-day attacks, while the training process is resistant to SPoF, data tampering, and up to 33% poisoning nodes. Performance evaluation illustrates the efficiency of MetaCIDS with 96% to 99% detection accuracy on four different network intrusion datasets, supporting both multi-class detection using labeled data and anomaly detection trained on unlabeled data.

A common problem in medical applications is predicting the time until an event of interest such as the onset of a disease, time to tumor recurrence, and time to mortality. Traditionally, classical survival analysis techniques have been used to address this problem. However, these techniques are of limited usage when considering nonlinear and interaction effects among biomarkers, and high profiling survival datasets. Although supervised machine learning techniques have shown some advantages over standard statistical methods in handling high-dimensional datasets, their application to survival analysis, particularly in the context of feature-based approaches, is at best limited. A major reason behind this is the difficulty in processing censored data, which is a common component of survival analysis. In this paper, we have transformed the time-to-event prediction problem into a semi-supervised regression problem. We utilize a self-training wrapper approach, where an outer layer guides the iterative refinement of predictions. This approach enhances the performance of our model by leveraging confident predictions from censored instances. The self-training wrapper is applied in conjunction with random survival forests as the base learner. In this approach, censored observations are introduced as partially labeled observations since their predicted time (target value) should exceed the censoring time. First, the algorithm builds a base model over the observed instances and then augments them iteratively with highly confident predictions over the censored set, using a smart stopping criterion based on the censoring time. The proposed approach has been evaluated and compared on fifteen real-world survival analysis datasets, including clinical and high-dimensional data. The ability of our proposed approach to integrate partial supervision information within a semi-supervised learning strategy has enabled it to achieve competitive performance compared to baseline models, particularly in the case of a high-dimensional regime.

Given Natural Language (NL) text descriptions, NL-based vehicle retrieval aims to extract target vehicles from a multi-view multi-camera traffic video pool. Solutions to the problem have been challenged by not only inherent distinctions between textual and visual domains, but also by the complexities of the high-dimensionality of visual data, the diverse range of textual descriptions, a major lack of high-volume datasets in this relatively new field, alongside prominently large domain gaps between training and test sets. To deal with these issues, existing approaches have advocated computationally expensive models to separately extract the subspaces of language and vision before blending them into the same shared representation space. Through our proposed Domain Adaptive Knowledge-based Retrieval System (DAKRS), we show that by taking advantage of multi-modal information in a pretrained model, we can better focus on training robust representations in the shared space of limited labels, rather than on robust extraction of uni-modal representations that comes with increased computational burdens. Our contributions are threefold: (i) An efficient extension of Contrastive Language-Image Pre-training (CLIP)’s transfer learning into a baseline text-to-image multi-modular vehicle retrieval framework; (ii) A data enhancement method to create pseudo-vehicle tracks from the traffic video pool by leveraging the robustness of baseline retrieval model combined with background subtraction; and (iii) A Semi-Supervised Domain Adaptation (SSDA) scheme to engineer pseudo-labels for adapting model parameters to the target domain. Experimental results are benchmarked on Cityflow-NL to obtain 63.20% MRR with 150.0 M of parameters, illustrating our competitive effectiveness and efficiency against state-of-the-arts, without ensembling.

Missing data imputation is a technique to deal with incomplete datasets. Since many models and algorithms cannot be applied to data containing missing values, a pre-processing step needs to be performed to remove incomplete data or to estimate the missing values. This is a well-known problem referred to as the data imputation problem. Several approaches have been designed for data imputation. These algorithms can be divided into two main categories: statistical and machine learning-based algorithms. As machine learning algorithms are optimized, they usually have better performance compared with statistical ones. In this paper, we review the most recent literature related to missing data imputation based on generative adversarial networks (GANs) that have gained tremendous attention in dealing with missing values. We examine the structures of GANs for missing data imputation and discuss the commonly used datasets and metrics for evaluation. We also cover the influence of the missing datatype, the effect of the missing data fraction, and the algorithm-related problems on data imputation performance. We conduct experiments on two publicly available datasets and evaluate the performance of GAIN, a missing data imputation algorithm to that of existing state-of-the-art approaches, demonstrating that the GAN-based algorithm outperforms the others in terms of RMSE and FID.

Metallographic image segmentation is a core task towards the automation of metallographic analysis. Currently, the most advanced methods for this task generally employ supervised deep learning segmentation models that require a great number of pixel-level annotated images, while the annotation process is time-consuming and labor-intensive. In order to address this issue, a semi-supervised model called Con2Net is proposed in this work, which leverages unlabeled data to improve performance for metallographic image segmentation. The Con2Net model adopts a multi-decoder architecture, which enforces consistency constraint between each decoder’s output and other decoders’ soft pseudo labels produced by output sharpening. In addition, to mitigate the negative impact caused by sharpening on false predicted pixels, we adopt the sharpening operation only to accurately predicted pixels. For labeled pixels, comparing ground truth to filter out correctly predicted pixels is the simplest and most effective approach. For unlabeled pixels, a contrastive learning module is introduced, which encourages the model to have better intra-class compactness and inter-class dispersion in the feature space. Based on that, pseudo-labels for unlabeled pixels are obtained by calculating the maximum similarity of feature vectors, and then the accurately predicted unlabeled pixels could be filtered out. We conduct experiments on two public datasets for metallographic image segmentation, comparing the proposed Con2Net model with five state-of-the-art semi-supervised segmentation models on three semi-supervised data partition protocols. The results demonstrate that Con2Net not only outperforms the supervised baseline by a significant margin, but also achieves superior segmentation performance compared to other five semi-supervised models. Our source code is available at https://github.com/Siiimon2423/Con2Net.

Load forecasting of shore power (SP) plays an important role in the power decision-making of the electrical grid due to docked ships are necessary to plug into the electrical grid. However, obtaining a large amount of labeled data on docked ships is time-consuming, presenting a challenge for Shore Power Load Forecasting. Additionally, multiple raw information entries for docked ships could lead to feature redundancy. To address these issues, we proposed a novel three-stage load forecasting method which includes attributive feature selection, semi-supervised learning (SSL) method for the mean of load distribution prediction, and a transformer-based model for variance prediction. Firstly, Firefly Algorithm (FA) is adopted to extract representative attribute features of docked ships to deal with the feature redundancy. Next, the selected feature set and label set are divided into two parts: a few labeled data and a large amount of labeled data. And we propose a $\Pi $ -model-based SSL method to predict the load distribution. Finally, we propose a transformer-based model to predict the variance of load distribution. Our model takes into account all historical load data of each docked ship for context learning. Further, we consider that the attribute features would also affect the variance prediction, so the latent features of the $\Pi $ -model are served as the initial condition which concatenates historical load data. We evaluated our model using 328 power load data from various ships that berth at Zhenjiang Port with shore power, totaling approximately 21,521 hours. The experiments prove the accuracy and efficiency of our proposed model, producing promising forecasting results.

Surface electromyography (sEMG) based gesture recognition has received broad attention and application in rehabilitation areas for its direct and fine-grained sensing ability. sEMG signals exhibit strong user dependence properties among users with different physiology, causing the inapplicability of the recognition model on new users. Domain adaptation is the most representative method to reduce the user gap with feature decoupling to acquire motion-related features. However, the existing domain adaptation method shows awful decoupling results when handling complex time-series physiological signals. Therefore, this paper proposes an Iterative Self-Training based Domain Adaptation method (STDA) to supervise the feature decoupling process with the pseudo-label generated by self-training and to explore cross-user sEMG gesture recognition. STDA mainly consists of two parts, discrepancy-based domain adaptation (DDA) and pseudo-label iterative update (PIU). DDA aligns existing users’ data and new users’ unlabeled data with a Gaussian kernel-based distance constraint. PIU Iteratively continuously updates pseudo-labels to generate more accurate labelled data on new users with category balance. Detailed experiments are performed on publicly available benchmark datasets, including the NinaPro dataset (DB-1 and DB-5) and the CapgMyo dataset (DB-a, DB-b, and DB-c). Experimental results show that the proposed method achieves significant performance improvement compared with existing sEMG gesture recognition and domain adaption methods.

The self-training methods have been praised by extensive research in semi-supervised classification. Mislabeling is the main challenge in self-training methods. Multiple variations of self-training methods are recently proposed against mislabeling from the following one of two aspects: a) using heuristic rules to find high-confidence unlabeled samples that can easily be predicted correctly in each iteration; b) enhancing prediction performance by employing ensemble classifiers composed of multiple weak classifiers. Yet, they still suffer from the following issues: a); most strategies for finding high-confidence unlabeled samples heavily rely on parameters; b) almost all employed ensemble classifiers originally designed for supervised classifiers and may not be suitable for semi-supervised classification due to the limited number and unrepresented distribution of the initial labeled data; c) few can overcome mislabeling from the above two aspects at the same time. To advance the state of the art, a new self-training method based on density peaks clustering and improved Adaboost is presented and named as STDPboost. In the iterative self-taught process, a new density peaks clustering-based strategy is proposed to find high-confidence unlabeled samples and a new ensemble classifier named AdaboostSEMI and more suitable for semi-supervised classification is proposed to predict high-confidence unlabeled samples, which overcomes mislabeling and the mentioned shortcomings of existing self-training methods. Intensive experiments on benchmark data sets have proven that STDPboost outperforms 7 state-of-the-art self-training methods in average classification accuracy of KNN classifier and CART classifier with the percentages of the initial labeled data from 10% to 50% due to further alleviating mislabeling.

A large number of methods for anomaly detection in robotic manipulation have been proposed, but their applicability and performance in real-world scenarios are often not established. In this paper, we therefore perform an experimental comparison of a broad range of practically applicable methods to detect exogenous anomalies in pick-and-place tasks. We first collect a dataset, which has been made freely available, on a state-of-the-art collaborative robot. The extensive experimental campaign comprises 600 runs under normal operation, and 80 runs where exogenous perturbations are present, and data is collected on the robot’s joints, force-torque readings, and the time spent in each program node. Using the data collected during normal operation, we train a set of anomaly detection methods whose computational complexity is low enough to run on resource-constrained robot hardware. We then evaluate the trained methods on the data collected in the perturbed runs and find that several methods can achieve a high anomaly detection performance. We show that exploiting knowledge of the robot’s program tree can increase the performance for some types of anomalies. We also observe that performance, in general, deteriorates when the application includes: (i) rarely-visited program branches, (ii) physical contact with the environment, and (iii) stochastic trajectories.

Nowadays, most people obtain information from social media networks, where news accompanied by photos and videos attracts readers more than traditional ones. However, these advantages are often misused by some publishers to disseminate fake information rapidly, thereby adversely affecting individuals and societies. Thus, the early detection of fake posts is crucial. Developing an automatic content-based fake news detector is the ideal way to overcome this issue. Given that the generation rate of news in social media is drastic and the labeling of a huge amount of data required by fully supervised models is expensive and time consuming, these models are not beneficial in real applications. To address this limitation, this study presents a semi-supervised method by utilizing an ensemble of diverse deep learners, effective augmentations, and a distribution-aware pseudo-labeling technique. Here, the proposed hybrid loss function enforces the learners to have accurate classification performance while attending to different parts of news content. Moreover, the proposed augmentations enhance the robustness of learners and prevent overfitting effectively. Diverse learners are utilized to annotate the unlabeled posts accurately and update their parameters from the most confident predicted news in a curriculum way, thereby enhancing the quality of pseudo labels and the robustness of the model. Moreover, we utilize encoded sentences from pre-trained transformer models, such as XLNET, and parameter sharing to build light learners on a common deep feature extractor module. Consequently, while the number of parameters is less than that of the existing methods, experiments conducted on three public fake news datasets reveal that the proposed method consistently outperforms state-of-the-art models with different proportions of labeled data across all evaluated datasets.

To provide accurate key variable prediction, data-driven soft sensing techniques have extracted much attention in recent years. Due to different control strategies in industrial processes, it is noticed that the variables in the control loops can be autocorrelated while the others may be static, which needs to be considered simultaneously. In this paper, a quality-related concurrent dual-latent variable (CDLV) model is proposed for soft sensing construction. Two different kinds of latent variables are adopted to learn quality-related dynamic information and quality-related static information respectively. Both quality-related variables are then applied for quality prediction purposes. On this basis, the CDLV model is extended to a semi-supervised form to provide a comprehensive description for the soft sensor design with insufficient quality information. The proposed models are demonstrated by two industrial cases which show superiority over other relative methods in the accuracy of key quality variables prediction.

Self-supervised learning for semi-supervised facial expression recognition aims to avoid the need to collect expensive labeled facial expression data. Existing methods demonstrate an impressive performance boost, but they artificially assume that small labeled facial expression data and large unlabeled facial expression data are from the same class distribution. In a more realistic scenario, when utilizing facial expression data from a large face recognition database as unlabeled data, there will be a mismatch distribution between the two sets of data. This often results in severe performance degradation due to incorrect propagation of unlabeled data from unrelated sources. In this paper, we address the class distribution mismatch problem in deep semi-supervised learning-based facial expression recognition. Specifically, we propose a silhouette coefficient-based contrast clustering algorithm, which determines the degree of separation between clusters by examining the intra-cluster and inter-cluster distances to accurately detect out-of-distribution data. In addition, we propose a pseudo-labeling rethinking strategy that matches the soft pseudo-labels estimated from a fine-tuned network to the contrast clustering to produce reliable pseudo-labels. Experiments on three in-the-wild datasets, RAF-DB, FERPlus and AffectNet, demonstrate the effectiveness of our method, and our approach performs well compared to state-of-the-art methods.