Search

Engineering materials in constructed systems in service exhibit complex patterns, including structural damage, environmental artifacts, and artificial anomalies. In recent years, machine vision methods have been extensively studied, most of which train models using regular grey or color images in the visible bands and label at pixel levels with a large volume of data. The authors propose using hyperspectral imaging (HSI) for structural material condition assessment in this work. Compared with visible images, the research challenge is that HSI pixels with high-dimensional spectral profiles are beyond human perceptive capabilities with hidden discriminative power. Learning from labeled and unlabeled data is one direct approach to unlocking this power. A deep neural network-enabled spatial-spectral feature extraction and a semisupervised learning architecture were developed in this work. A human-in-the-loop (HITL) framework was comparatively studied with three incremental training-data configuration schemes. The paper concludes with the following empirical findings: (1) fully supervised learning determines the baseline of the detection performance; (2) an extensive range of ratio values exists between the unlabeled and the labeled data for incremental semisupervised learning, and a 1∶1 ratio can be taken as a conservative and operational ratio; and (3) with parametric semisupervised learning with equal labeled and unlabeled data participation, the proposed HITL operational workflow can be implemented as a practical approach for HSI-based structural material and damage detection. [ABSTRACT FROM AUTHOR]

Deep learning has emerged as a powerful tool to model complicated relationships between inputs and outputs in manufacturing systems. Existing deep learning approaches in manufacturing are often used to directly predict the Variables of Interest (VoI) such as the system status from a set of sensor measurements by supervised learning. However, in various complex manufacturing systems, components are operated under multiple modes that are not well known beforehand. The mapping of the VoI from sensor measurements highly depends on the mode information given that sensor measurements under different operation modes usually present different patterns. Therefore, predicting the VoI under multiple operation modes given sensor measurements is urgently necessary. This study develops a novel deep learning method for multimodal regression and mode recognition to predict the VoI under multiple modes and recognize the specific mode of a component from its sensor measurements. Specifically, we establish a deep neural network (DNN)-based regression- and classification-integrated framework. For model training, our innovative idea is to develop an Expectation–Maximum (EM)-based backpropagation algorithm, where the modes of components are set as latent variables, given that the mode information cannot be provided. Numerical experiments and a case study of degraded gas turbine engines are presented to validate the proposed model performance. [ABSTRACT FROM AUTHOR]

In deep mineral exploration, it is difficult to constrain the complex geological structures using a single geophysical method. To tackle the difficulty, integrated geophysical surveys and joint data interpretation are essential. Machine learning (ML) provides more accurate predictions than traditional methods, especially when dealing with complex data from multiple sources or varied statistical distributions. However, a major challenge in using ML for deep mineral exploration is the scarcity and imbalance of labeled samples, mainly due to budget constraints and the complexity of ore deposits. This issue reduces the accuracy of predictive models and introduces bias. Additionally, limited labeling can lead to difficulties in predicting previously undefined classes in training datasets. To address these challenges, we introduce a robust semisupervised ML framework that integrates diverse geophysical and geological datasets to improve model reliability with limited labeled data. Our approach uses a semisupervised ML variational Gaussian mixture model (SsL-VGMM) to handle issues related to insufficient and imbalanced data. We enhanced the model's predictive capability for unseen data by introducing a novel penalty factor in the 'cannot-link' function. Moreover, we employed Bayesian optimization, focusing on the mean-mixture weight, to avoid local optima during model training. Our model demonstrated high accuracy and efficiency, with classification and prediction accuracies of 95.33% and 87.4%, respectively, in numerical and electromagnetic simulation scenarios. Its effectiveness was further validated by locating Pb–Zn–Ag deposits in Inner Mongolia, supported by actual drilling data. This paper highlights the model's potential in complex mineral exploration and its significant practical and innovative value for deep mineral exploration. [ABSTRACT FROM AUTHOR]

This paper presents Fix-YOLOX (Fixmatch-You Only Look Once X), a semisupervised target detection model that uses a small amount of annotated data for fully supervised training, and adds a semisupervised training module using both pseudolabelling and consistent regularization to prevent overfitting in fully supervised training by using unlabelled data. Additionally, the generalization of the model and its fault tolerance to labelled data are improved. The experimental results show that the proposed semisupervised visual detection algorithm, Fix-YOLOX, can substantially reduce the amount of data annotation required for the target detection task while effectively overcoming the problem caused by annotated data with uneven quality. The YOLOX model achieves 91.95% accuracy with 50% annotated data and an average detection time of 10.4 ms per image/frame, which is consistent with the detection time of original YOLOX. Therefore, the model has good real-time performance and generalizability. [ABSTRACT FROM AUTHOR]

Obstructive sleep apnea (OSA) is a prevalent sleep disorder that seriously affects patients' quality of life and health status. Traditional diagnostic methods are time-consuming and laborintensive, and apnea detection using deep learning algorithms also faces the problems of insufficient sample size and class imbalance. Therefore, in this paper, we propose a semisupervised apnea detection algorithm (Semi-DynaSeqNet) based on the oxygen saturation (SpO2) signal. In this study, we first extracted local features of the SpO2 signal using a onedimensional convolutional neural network, and then combined the gate recurrent unit for time series modeling to capture the signal's long-term dynamic features. On this basis, a self-attention mechanism is introduced to further enhance the recognition of key features. Considering the small-sample classification task of the OSA detection, we further proposed the semisupervised learning method with the adaptive adjustment of threshold. By iteratively training a model to generate pseudo-labeled samples of unlabeled pulse oximetry signals and incorporating them into the training set, while adaptively adjusting the semisupervised threshold to fully utilize the unlabeled sample information, we thereby improved the generalization ability of Semi-DynaSeqNet. The experimental results showed that the algorithm proposed in this paper achieves an F1-score of 90.94% for the model on the St. Vincent's University Hospital/University College Dublin Sleep Apnea Database with one-second detection, whereas the algorithm achieves an F1-score of 94.65% on the PhysioNet Apnea-ECG Database with one-minute detection, indicating that the algorithm can perform well in both sleep apnea detection tasks with different time scales, demonstrating its flexibility and scalability. [ABSTRACT FROM AUTHOR]

Laplace learning is a popular machine learning algorithm for finding missing labels from a small number of labeled feature vectors using the geometry of a graph. More precisely, Laplace learning is based on minimizing a graph-Dirichlet energy, equivalently a discrete Sobolev W2,1 seminorm, constrained to taking the values of known labels on a given subset. The variational problem is asymptotically ill-posed as the number of unlabeled feature vectors goes to infinity for finite given labels due to a lack of regularity in minimizers of the continuum Dirichlet energy in any dimension higher than one. In particular, continuum minimizers are not continuous. One solution is to consider higher-order regularization, which is the analogue of minimizing Sobolev Ws,2 seminorms. In this paper we consider the asymptotics of minimizing a graph variant of the Sobolev Ws,2 seminorm with pointwise constraints. We show that, as expected, one needs s > d/2, where d is the dimension of the data manifold. We also show that there must be an upper bound on the connectivity of the graph; that is, highly connected graphs lead to degenerate behavior of the minimizer even when s > d/2. [ABSTRACT FROM AUTHOR]

Abstract Globally, coal is a critical energy source, and the profits of related enterprises are highly related to changes in the coal price. A robust coal purchasing cost forecasting method may enhance the coal purchasing strategies of coal‐consuming enterprises and obtain key information for reducing global carbon emissions. However, forecasting the price of coal is a challenging task due to the noise and high random fluctuation of coal price data. To overcome these obstacles, this research proposes a novel forecasting method combining data decomposition, semisupervised feature engineering, and ensemble learning to forecast coal prices. Initially, the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) method is employed to decompose the coal price series to reduce the complexity. Second, considering the fluctuation of coal price is influenced by various factors (such as transportation cost and coal mine production), the proposed system incorporates an adaptive data fusion module to fuse data from multiple sources. Finally, a stacking‐based ensemble learning model is adopted in the method to increase the forecasting accuracy by combining the forecasting results of multiple models. The Bohai‐Rim Steam‐Coal Price Index was used to validate the proposed method, and the result of the case study shows that the proposed method provides superior performance than the other nine baseline models in all measured indices. The outcomes of ablation tests indicate the precision of each algorithm is improved by combining CEEMDAN, which proves that the decomposition algorithm is necessary.

Training deep learning models for vision-based monitoring of construction sites usually requires a large amount of labeled data. Semisupervised learning methods can efficiently obtain unlabeled data with substantial cost savings. Thus, this paper proposes a semisupervised object detection method for construction site monitoring. Weather as well as strong and weak data augmentation are integrated to cope with the complex construction site conditions (weather changes, camera view shifts, and so on) by integrating semisupervised learning to leverage the valid information in unlabeled construction site images. To validate its effectiveness, the proposed method was tested on the Alberta Construction Image Data Set (ACID), a public data set for the construction research community. The experimental results revealed that the proposed method achieves an average accuracy [mean average precision (mAP)] of 81.1% when trained on only 3% of the labeled images. This study helps to significantly reduce the development cost of vision-based object detection models for construction sites. [ABSTRACT FROM AUTHOR]

Globally, coal is a critical energy source, and the profits of related enterprises are highly related to changes in the coal price. A robust coal purchasing cost forecasting method may enhance the coal purchasing strategies of coal‐consuming enterprises and obtain key information for reducing global carbon emissions. However, forecasting the price of coal is a challenging task due to the noise and high random fluctuation of coal price data. To overcome these obstacles, this research proposes a novel forecasting method combining data decomposition, semisupervised feature engineering, and ensemble learning to forecast coal prices. Initially, the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) method is employed to decompose the coal price series to reduce the complexity. Second, considering the fluctuation of coal price is influenced by various factors (such as transportation cost and coal mine production), the proposed system incorporates an adaptive data fusion module to fuse data from multiple sources. Finally, a stacking‐based ensemble learning model is adopted in the method to increase the forecasting accuracy by combining the forecasting results of multiple models. The Bohai‐Rim Steam‐Coal Price Index was used to validate the proposed method, and the result of the case study shows that the proposed method provides superior performance than the other nine baseline models in all measured indices. The outcomes of ablation tests indicate the precision of each algorithm is improved by combining CEEMDAN, which proves that the decomposition algorithm is necessary. [ABSTRACT FROM AUTHOR]

In many modern machine learning applications, changes in covariate distributions and difficulty in acquiring outcome information have posed challenges to robust model training and evaluation. Numerous transfer learning methods have been developed to robustly adapt the model itself to some unlabeled target populations using existing labeled data in a source population. However, there is a paucity of literature on transferring performance metrics, especially receiver operating characteristic (ROC) parameters, of a trained model. In this paper, we aim to evaluate the performance of a trained binary classifier on unlabeled target population based on ROC analysis. We proposed S emisupervised T ransfer l E arning of A ccuracy M easures (STEAM), an efficient three-step estimation procedure that employs (1) double-index modeling to construct calibrated density ratio weights and (2) robust imputation to leverage the large amount of unlabeled data to improve estimation efficiency. We establish the consistency and asymptotic normality of the proposed estimator under the correct specification of either the density ratio model or the outcome model. We also correct for potential overfitting bias in the estimators in finite samples with cross-validation. We compare our proposed estimators to existing methods and show reductions in bias and gains in efficiency through simulations. We illustrate the practical utility of the proposed method on evaluating prediction performance of a phenotyping model for rheumatoid arthritis (RA) on a temporally evolving EHR cohort. [ABSTRACT FROM AUTHOR]

Molecular prediction tasks normally demand a series of professional experiments to label the target molecule, which suffers from the limited labeled data problem. One of the semisupervised learning paradigms, known as self-training, utilizes both labeled and unlabeled data. Specifically, a teacher model is trained using labeled data and produces pseudo labels for unlabeled data. These labeled and pseudo-labeled data are then jointly used to train a student model. However, the pseudo labels generated from the teacher model are generally not sufficiently accurate. Thus, we propose a robust self-training strategy by exploring robust loss function to handle such noisy labels in two paradigms, that is, generic and adaptive. We have conducted experiments on three molecular biology prediction tasks with four backbone models to gradually evaluate the performance of the proposed robust self-training strategy. The results demonstrate that the proposed method enhances prediction performance across all tasks, notably within molecular regression tasks, where there has been an average enhancement of 41.5%. Furthermore, the visualization analysis confirms the superiority of our method. Our proposed robust self-training is a simple yet effective strategy that efficiently improves molecular biology prediction performance. It tackles the labeled data insufficient issue in molecular biology by taking advantage of both labeled and unlabeled data. Moreover, it can be easily embedded with any prediction task, which serves as a universal approach for the bioinformatics community. [ABSTRACT FROM AUTHOR]

The development of speech technology requires large amounts of data to estimate the underlying models. Even when relying on large multilingual pre-trained models, some amount of task-specific data on the target language is needed to fine-tune those models and obtain competitive performance. In this paper, we present a bilingual Basque–Spanish dataset extracted from parliamentary sessions. The dataset is designed to develop and evaluate automatic speech recognition (ASR) systems but can be easily repurposed for other speech-processing tasks (such as speaker or language recognition). The paper first compares the two target languages, emphasizing their similarities at the acoustic-phonetic level, which sets the basis for sharing data and compensating for the relatively small amount of spoken resources available for Basque. Then, Basque Parliament plenary sessions are characterized in terms of organization, topics, speaker turns and the use of the two languages. The paper continues with the description of the data collection procedure (involving both speech and text), the audio formats and conversions along with the creation and postprocessing of text transcriptions and session minutes. Then, it describes the semi-supervised iterative procedure used to cut, rank and select the training segments and the manual supervision process employed to produce the test set. Finally, ASR experiments are presented using state-of-the-art technology to validate the dataset and to set a reference for future works. The datasets, along with models and recipes to reproduce the experiments reported in the paper, are released through Hugging Face. [ABSTRACT FROM AUTHOR]

Electroencephalogram (EEG) data annotation demands considerable expertise and is a time-intensive process. Moreover, inter-subject variability intensifies the challenge of domain shift, adversely impacting the generalization performance of deep learning models on unseen subjects. Current methods in EEG data analysis often struggle to handle the complex nature of brain activity without relying on EEG feature engineering. In this paper, we present a hybrid semi-supervised framework for seizure type classification, which relies on minimal domain knowledge provided by exploiting spectral and spatial patch-level representations of raw unlabeled EEG data, while leveraging a small amount of labeled data. Our method, IMPRESS, enhances EEG representation learning by combining multi-patch mutual information maximization with adversarial distribution alignment. We assessed the framework’s performance for cross-patient seizure classification using publicly accessible Temple University Seizure Corpus. IMPRESS surpasses the best-performing semi-supervised learning method by 1.92% and 0.72% using balanced accuracy and macro-F1 metrics, respectively, with 40 labeled samples per class. Remarkably, IMPRESS surpasses the fully-supervised method while requiring only 25 labeled samples per class. Additionally, we visualize the learned feature embeddings, highlighting the underlying dynamics across different seizure types, aiding in understanding the model’s behavior. This demonstrates the potential of leveraging multi-patch information from unlabeled data through a contrastive data-driven approach, alleviating the burden of annotating large amounts of EEG data.

Hyperspectral images are valuable for precise land cover change detection in a consistent area over time. Nevertheless, supervised methods for hyperspectral change detection face limitations due to insufficient labeled samples. Additionally, current deep learning-based approaches often neglect critical feature interactions across temporal, spatial, and spectral domains. To address these challenges, we introduce a semisupervised model called the HyperMatch-based Cross Temporal and Spatial Interaction Transformer (CTSIT) for hyperspectral change detection. The key contributions of this study are as follows: Introduction of the HyperMatch training schedule, which relies on weak-to-strong consistency, to enhance feature extraction using both labeled and unlabeled data. Then, the contribution of the Cross Temporal Bidirectional Attention module to emphasize bidirectional temporal interactions and resemblances. Finally, Introduction of the Spatial and Spectral Attention module, which includes the Dense Cross Spatial Attention and Spectral Attention modules, to capture long-range densely spatial interactions and internal spectral similarities. Extensive comparative experiments conducted on three mainstream hyperspectral change detection datasets confirm the effectiveness and superiority of the proposed HyperMatch-based CTSIT in utilizing both labeled and unlabeled samples. For instance, the training ablation demonstrates that this method significantly outperforms most existing state-of-the-art hyperspectral change detection methods, even with a significantly smaller number of samples.

Synthetic aperture radar (SAR) image target detection methods based on semisupervised learning, such as the mean teacher framework, have shown promise in diminishing the issue of limited labeled data. However, several challenges exist in current methods. First, data augmentation techniques designed for optical images may not be suitable for SAR images due to differences in imaging methods. In addition, the contribution of pseudo labels remains constant during the initial retraining stage can lead to degradation in prediction results. Moreover, the low quality of predicted bounding boxes poses a challenge for effective retraining. To address these challenges, we propose an end-to-end semisupervised detection method based on the mean teacher framework. To enhance the robustness of training, we first introduce SAR-specific data augmentation techniques, including multiplicative noise, which effectively increase the diversity of training samples. Second, we propose a method that weights the losses of pseudo-labeled data using a hard-sigmoid function, gradually improving the importance of pseudo-labeled data during retraining, thereby alleviating their potential negative impact on the training process. Finally, we propose an IoU-aware subnetwork to incorporate high-quality pseudo-labeled bounding boxes into retraining, allowing them to contribute to network adjustments while mitigating the impact of low-quality samples. Experimental evaluations on publicly available SAR image datasets demonstrate the effectiveness of our proposed method in improving the target detection capability of semisupervised SAR target detection.

Stroke is the leading cause of permanent disability in adults, and it can cause permanent brain damage. According to the World Health Organization, 795,000 Americans experience a new or recurrent stroke each year. Early detection of medical disorders, for example, strokes, can minimize the disabling effects. Thus, in this paper, we consider various risk factors that contribute to the occurrence of stoke and machine learning algorithms, for example, the decision tree, random forest, and naive Bayes algorithms, on patient characteristics survey data to achieve high prediction accuracy. We also consider the semisupervised self-training technique to predict the risk of stroke. We then consider the near-miss undersampling technique, which can select only instances in larger classes with the smaller class instances. Experimental results demonstrate that the proposed method obtains an accuracy of approximately 98.83% at low cost, which is significantly higher and more reliable compared with the compared techniques.

Abstract Most unsupervised or semisupervised hyperspectral anomaly detection (HAD) methods train background reconstruction models in the original spectral domain. However, due to the noise and spatial resolution limitations, there may be a lack of discrimination between backgrounds and anomalies. This makes it easy for the autoencoder to capture the low‐level features shared between the two, thereby increasing the difficulty of separating anomalies from the backgrounds, which runs counter to the purpose of HAD. To this end, the authors map the original spectrums to the fractional Fourier domain (FrFD) and reformulate it as a mapping task in which restoration errors are employed to distinguish background and anomaly. This study proposes a novel frequency‐to‐spectrum mapping generative adversarial network for HAD. Specifically, the depth separable features of backgrounds and anomalies are enhanced in the FrFD. Due to the semisupervised approach, FTSGAN needs to learn the embedded features of the backgrounds, thus mapping and restoring them from the FrFD to the original spectral domain. This strategy effectively prevents the model from focussing on the numerical equivalence of input and output, and restricts the ability of FTSGAN to restore anomalies. The comparison and analysis of the experiments verify that the proposed method is competitive.

A novel approach for video instance segmentation is presented using semisupervised learning. Our Cluster2Former model leverages scribble-based annotations for training, significantly reducing the need for comprehensive pixel-level masks. We augment a video instance segmenter, for example, the Mask2Former architecture, with similarity-based constraint loss to handle partial annotations efficiently. We demonstrate that despite using lightweight annotations (using only 0.5% of the annotated pixels), Cluster2Former achieves competitive performance on standard benchmarks. The approach offers a cost-effective and computationally efficient solution for video instance segmentation, especially in scenarios with limited annotation resources. [ABSTRACT FROM AUTHOR]

NONNEGATIVE matrix factorization (NMF) is an effective technique for dimensionality reduction of high-dimensional data for tasks such as machine learning and data visualization. However, for practical clustering tasks, traditional NMF ignores the manifold information of both the data space and feature space, as well as the discriminative information of the data. In this paper, we propose a semisupervised NMF called dual-graph-regularization-constrained nonnegative matrix factorization with label discrimination (DCNMFLD). DCNMFLD combines dual graph regularization and prior label information as additional constraints, making full use of the intrinsic geometric and discriminative structures of the data, and can efficiently enhance the discriminative and exclusionary nature of clustering and improve the clustering performance. The evaluation of the clustering experimental results on four benchmark datasets demonstrates the effectiveness of our new algorithm. [ABSTRACT FROM AUTHOR]

Purpose: In this paper, we propose a method to classify and detect normal, known anomalies, and unknown anomalies by combining the expectation-maximisation (EM-λ) algorithm and the Mahalanobis-Taguchi (MT) method. Methodology/Approach: The proposed method learns normal data that are expected to be homogeneous and known abnormal data and performs classification and detection by parameter estimation using the EM-λ algorithm. Conventional methods perform analysis based on parameter estimation using the EM algorithm. However, the EM algorithm can degrade classification accuracy if it does not assume that the data fits the model's generative process. Findings: We verify the performance of the proposed method using artificially generated data and real-world bean data for classification as data that do not satisfy this assumption. The validation results show up to 6% improvement over the conventional method in classification accuracy and unknown anomaly discrimination accuracy. Research Limitation/implication: We try various patterns for the parameter of the proposed method in the verification. However, this way is computationally expensive. Originality/Value of paper: Conventional methods perform analysis based on parameter estimation using the EM algorithm. Our proposal method seeks to improve accuracy by using the EM-λ algorithm for parameter estimation, which is expected to improve classification accuracy when the data do not conform to the generative assumptions of the EM algorithm's model. Category: Conceptual paper. [ABSTRACT FROM AUTHOR]

Stroke is the leading cause of permanent disability in adults, and it can cause permanent brain damage. According to the World Health Organization, 795 000 Americans experience a new or recurrent stroke each year. Early detection of medical disorders, for example, strokes, can minimize the disabling effects. Thus, in this paper, we consider various risk factors that contribute to the occurrence of stoke and machine learning algorithms, for example, the decision tree, random forest, and naive Bayes algorithms, on patient characteristics survey data to achieve high prediction accuracy. We also consider the semisupervised self‐training technique to predict the risk of stroke. We then consider the near‐miss undersampling technique, which can select only instances in larger classes with the smaller class instances. Experimental results demonstrate that the proposed method obtains an accuracy of approximately 98.83% at low cost, which is significantly higher and more reliable compared with the compared techniques. [ABSTRACT FROM AUTHOR]

Most unsupervised or semisupervised hyperspectral anomaly detection (HAD) methods train background reconstruction models in the original spectral domain. However, due to the noise and spatial resolution limitations, there may be a lack of discrimination between backgrounds and anomalies. This makes it easy for the autoencoder to capture the low‐level features shared between the two, thereby increasing the difficulty of separating anomalies from the backgrounds, which runs counter to the purpose of HAD. To this end, the authors map the original spectrums to the fractional Fourier domain (FrFD) and reformulate it as a mapping task in which restoration errors are employed to distinguish background and anomaly. This study proposes a novel frequency‐to‐spectrum mapping generative adversarial network for HAD. Specifically, the depth separable features of backgrounds and anomalies are enhanced in the FrFD. Due to the semisupervised approach, FTSGAN needs to learn the embedded features of the backgrounds, thus mapping and restoring them from the FrFD to the original spectral domain. This strategy effectively prevents the model from focussing on the numerical equivalence of input and output, and restricts the ability of FTSGAN to restore anomalies. The comparison and analysis of the experiments verify that the proposed method is competitive. [ABSTRACT FROM AUTHOR]

PurposeTo develop a convolutional neural network (CNN)-based deformable lung registration algorithm to reduce computation time and assess its potential for lobar air trapping quantification.Materials and methodsIn this retrospective study, a CNN algorithm was developed to perform deformable registration of lung CT (LungReg) using data on 9118 patients from the COPDGene Study (data collected between 2007 and 2012). Loss function constraints included cross-correlation, displacement field regularization, lobar segmentation overlap, and the Jacobian determinant. LungReg was compared with a standard diffeomorphic registration (SyN) for lobar Dice overlap, percentage voxels with nonpositive Jacobian determinants, and inference runtime using paired t tests. Landmark colocalization error (LCE) across 10 patients was compared using a random effects model. Agreement between LungReg and SyN air trapping measurements was assessed using intraclass correlation coefficient. The ability of LungReg versus SyN emphysema and air trapping measurements to predict Global Initiative for Chronic Obstructive Lung Disease (GOLD) stages was compared using area under the receiver operating characteristic curves.ResultsAverage performance of LungReg versus SyN showed lobar Dice overlap score of 0.91-0.97 versus 0.89-0.95, respectively (P < .001); percentage voxels with nonpositive Jacobian determinant of 0.04 versus 0.10, respectively (P < .001); inference run time of 0.99 second (graphics processing unit) and 2.27 seconds (central processing unit) versus 418.46 seconds (central processing unit) (P < .001); and LCE of 7.21 mm versus 6.93 mm (P < .001). LungReg and SyN whole-lung and lobar air trapping measurements achieved excellent agreement (intraclass correlation coefficients > 0.98). LungReg versus SyN area under the receiver operating characteristic curves for predicting GOLD stage were not statistically different (range, 0.88-0.95 vs 0.88-0.95, respectively; P = .31-.95).ConclusionCNN-based deformable lung registration is accurate and fully automated, with runtime feasible for clinical lobar air trapping quantification, and has potential to improve diagnosis of small airway diseases.Keywords: Air Trapping, Convolutional Neural Network, Deformable Registration, Small Airway Disease, CT, Lung, Semisupervised Learning, Unsupervised Learning Supplemental material is available for this article. © RSNA, 2021 An earlier incorrect version of this article appeared online. This article was corrected on December 22, 2021.

A new paradigm for data-driven, model-agnostic new physics searches at colliders is emerging, and aims to leverage recent breakthroughs in anomaly detection and machine learning. In order to develop and benchmark new anomaly detection methods within this framework, it is essential to have standard datasets. To this end, we have created the LHC Olympics 2020, a community challenge accompanied by a set of simulated collider events. Participants in these Olympics have developed their methods using an R&D dataset and then tested them on black boxes: datasets with an unknown anomaly (or not). Methods made use of modern machine learning tools and were based on unsupervised learning (autoencoders, generative adversarial networks, normalizing flows), weakly supervised learning, and semi-supervised learning. This paper will review the LHC Olympics 2020 challenge, including an overview of the competition, a description of methods deployed in the competition, lessons learned from the experience, and implications for data analyses with future datasets as well as future colliders.

Images taken in various real-world scenarios meet the symmetrical goal of simultaneously removing foreground rain-induced occlusions and restoring the background details. This inspires us to remember the principle of symmetry; real-world rain is a mixture of rain streaks and rainy haze and degrades the visual quality of the background. Current efforts formulate image rain streak removal and rainy haze removal as separate models, which disrupts the symmetrical characteristics of real-world rain and background, leading to significant performance degradation. To achieve this symmetrical balance, we propose a novel semisupervised coarse-to-fine guided generative adversarial network (Semi-RainGAN) for the mixture of rain removal. Beyond existing wisdom, Semi-RainGAN is a joint learning paradigm of the mixture of rain removal and attention and depth estimation. Additionally, it introduces a coarse-to-fine guidance mechanism that effectively fuses estimated image, attention, and depth features. This mechanism enables us to achieve symmetrically high-quality rain removal while preserving fine-grained details. To bridge the gap between synthetic and real-world rain, Semi-RainGAN makes full use of unpaired real-world rainy and clean images, enhancing its generalization to real-world scenarios. Extensive experiments on both synthetic and real-world rain datasets demonstrate clear visual and numerical improvements of Semi-RainGAN over sixteen state-of-the-art models. [ABSTRACT FROM AUTHOR]

The training of supervised machine learning approaches is critically dependent on annotating large-scale datasets. Semisupervised learning approaches aim to achieve compatible performance with supervised methods using relatively less annotation without sacrificing good generalization capacity. In line with this objective, ways of leveraging unlabeled data have been the subject of intense research. However, semisupervised video action recognition has received relatively less attention compared to image domain implementations. Existing semisupervised video action recognition methods trained from scratch rely heavily on augmentation techniques, complex architectures, and/or the use of other modalities while distillation-based methods use models that have only been trained for 2D computer vision tasks. In another line of work, pretrained vision-language models have shown very promising results for generating general-purpose visual features with reports of high zero-shot performance for many downstream tasks. In this work, we exploit a language-supervised visual encoder for learning video representations for video action classification tasks. We propose a teacher-student learning paradigm through feature distillation and pseudo-labeling. Our experimental results are a proof-of-concept revealing that multimodal feature extractors can be utilized for spatiotemporal feature extraction in a semisupervised learning context and show compatible performance with SOTA methods, especially in a low-label regime. [ABSTRACT FROM AUTHOR]

A novel semisupervised hyperspectral imaging technique was developed to detect foreign materials (FMs) on raw poultry meat. Combining hyperspectral imaging and deep learning has shown promise in identifying food safety and quality attributes. However, the challenge lies in acquiring a large amount of accurately annotated/labeled data for model training. This paper proposes a novel semisupervised hyperspectral deep learning model based on a generative adversarial network, utilizing an improved 1D U-Net as its discriminator, to detect FMs on raw chicken breast fillets. The model was trained by using approximately 879,000 spectral responses from hyperspectral images of clean chicken breast fillets in the near-infrared wavelength range of 1000–1700 nm. Testing involved 30 different types of FMs commonly found in processing plants, prepared in two nominal sizes: 2 × 2 mm2 and 5 × 5 mm2. The FM-detection technique achieved impressive results at both the spectral pixel level and the foreign material object level. At the spectral pixel level, the model achieved a precision of 100%, a recall of over 93%, an F1 score of 96.8%, and a balanced accuracy of 96.9%. When combining the rich 1D spectral data with 2D spatial information, the FM-detection accuracy at the object level reached 96.5%. In summary, the impressive results obtained through this study demonstrate its effectiveness at accurately identifying and localizing FMs. Furthermore, the technique's potential for generalization and application to other agriculture and food-related domains highlights its broader significance. [ABSTRACT FROM AUTHOR]

Recent successes in point cloud semantic segmentation heavily rely on a large amount of annotated data. Furthermore, three-dimensional point cloud data are generally sparse and unorganized, and a frame of point cloud usually includes more than 100,000 points, which increases the difficulty of point cloud annotation. To reduce the annotation efforts, we propose a multi-granularity semisupervised active learning pipeline which aims to select representative, uncertain and diverse data to annotate. To better exploit annotating budget, we first leverage the conventional point cloud registration algorithm to develop a matching score function which is used to select a representative subset. And then we change the annotating units from a point cloud scan to segmented regions through two semisupervised methods. Subsequently, in each active selection step, segmented region information is calculated with two terms: softmax entropy and point cloud intensity, and the latter serves to encourage region diversity. Finally, to further reduce annotation effort, semisupervised learning is introduced to our pipeline to automatically select a portion of unlabeled segmented regions with high confidence and assign pseudolabels to them. Extensive experiments show that our approach greatly outperforms previous active learning methods, and we obtain the mean class intersection-over-union performance of 95% fully supervised learning with merely 3% of labeled data on SemanticKITTI dataset. [ABSTRACT FROM AUTHOR]

In this paper, a semisupervised speech data extraction method is presented and applied to create a new dataset designed for the development of fully bilingual Automatic Speech Recognition (ASR) systems for Basque and Spanish. The dataset is drawn from an extensive collection of Basque Parliament plenary sessions containing frequent code switchings. Since session minutes are not exact, only the most reliable speech segments are kept for training. To that end, we use phonetic similarity scores between nominal and recognized phone sequences. The process starts with baseline acoustic models trained on generic out-of-domain data, then iteratively updates the models with the extracted data and applies the updated models to refine the training dataset until the observed improvement between two iterations becomes small enough. A development dataset, involving five plenary sessions not used for training, has been manually audited for tuning and evaluation purposes. Cross-validation experiments (with 20 random partitions) have been carried out on the development dataset, using the baseline and the iteratively updated models. On average, Word Error Rate (WER) reduces from 16.57% (baseline) to 4.41% (first iteration) and further to 4.02% (second iteration), which corresponds to relative WER reductions of 73.4% and 8.8%, respectively. When considering only Basque segments, WER reduces on average from 16.57% (baseline) to 5.51% (first iteration) and further to 5.13% (second iteration), which corresponds to relative WER reductions of 66.7% and 6.9%, respectively. As a result of this work, a new bilingual Basque–Spanish resource has been produced based on Basque Parliament sessions, including 998 h of training data (audio segments + transcriptions), a development set (17 h long) designed for tuning and evaluation under a cross-validation scheme and a fully bilingual trigram language model. [ABSTRACT FROM AUTHOR]

Leveraging large-scale electronic health record (EHR) data to estimate survival curves for clinical events can enable more powerful risk estimation and comparative effectiveness research. However, use of EHR data is hindered by a lack of direct event time observations. Occurrence times of relevant diagnostic codes or target disease mentions in clinical notes are at best a good approximation of the true disease onset time. On the other hand, extracting precise information on the exact event time requires laborious manual chart review and is sometimes altogether infeasible due to a lack of detailed documentation. Current status labels—binary indicators of phenotype status during follow-up—are significantly more efficient and feasible to compile, enabling more precise survival curve estimation given limited resources. Existing survival analysis methods using current status labels focus almost entirely on supervised estimation, and naive incorporation of unlabeled data into these methods may lead to biased estimates. In this article, we propose Semisupervised Calibration of Risk with Noisy Event Times (SCORNET), which yields a consistent and efficient survival function estimator by leveraging a small set of current status labels and a large set of informative features. In addition to providing theoretical justification of SCORNET, we demonstrate in both simulation and real-world EHR settings that SCORNET achieves efficiency akin to the parametric Weibull regression model, while also exhibiting semi-nonparametric flexibility and relatively low empirical bias in a variety of generative settings. [ABSTRACT FROM AUTHOR]

The early screening and treatment of diabetic retinopathy (DR) and diabetic macular edema (DME) can prevent the risk of blindness for most diabetic patients. Joint grading of DR and DME can reduce the screening cost and improve screening efficiency. Most grading methods only focus on grading one disease and lack generality. In this paper, we propose an adaptive attention block (AAB) to better extract features and improve model performance, which can be adaptively adjusted according to different grading tasks. We propose the AABNet to handle multiple grading tasks, consisting of a lightweight architecture MobileNetv2 as a backbone for feature extraction and two independent branches, the DR branch and the DME branch, for grading prediction. To alleviate the lack of labeled data, we propose an adaptive teacher-student model as a semisupervised learning method to train the AABNet with additional unlabeled data, dependent on fewer model parameters, whose foundation is consistency regularization, which is added to the loss to provide additional supervised signals for training to obtain a more accurate model without destroying the image details and strictly requiring the quality of unlabeled images. Extensive experiments are conducted, including performance comparisons with state-of-the-art methods and ablation studies of attention mechanisms and semisupervised learning mechanisms. Experiments verify that AABNet can bring a satisfactory improvement in the multiple DR grading, DME grading and joint DR&DME grading tasks. AABNet outperforms other state-of-the-art methods and achieves better results in multiple grading tasks on the Messidor and DDR datasets. [ABSTRACT FROM AUTHOR]

Intrusion detection involves identifying unauthorized network activity and recognizing whether the data constitute an abnormal network transmission. Recent research has focused on using semi-supervised learning mechanisms to identify abnormal network traffic to deal with labeled and unlabeled data in the industry. However, real-time training and classifying network traffic pose challenges, as they can lead to the degradation of the overall dataset and difficulties preventing attacks. Additionally, existing semisupervised learning research might need to analyze the experimental results comprehensively. This paper proposes XA-GANomaly, a novel technique for explainable adaptive semi-supervised learning using GANomaly, an image anomalous detection model that dynamically trains small subsets to these issues. First, this research introduces a deep neural network (DNN)-based GANomaly for semi-supervised learning. Second, this paper presents the proposed adaptive algorithm for the DNN-based GANomaly, which is validated with four subsets of the adaptive dataset. Finally, this study demonstrates a monitoring system that incorporates three explainable techniques—Shapley additive explanations, reconstruction error visualization, and t-distributed stochastic neighbor embedding—to respond effectively to attacks on traffic data at each feature engineering stage, semi-supervised learning, and adaptive learning. Compared to other single-class classification techniques, the proposed DNN-based GANomaly achieves higher scores for Network Security Laboratory-Knowledge Discovery in Databases and UNSW-NB15 datasets at 13% and 8% of F1 scores and 4.17% and 11.51% for accuracy, respectively. Furthermore, experiments of the proposed adaptive learning reveal mostly improved results over the initial values. An analysis and monitoring system based on the combination of the three explainable methodologies is also described. Thus, the proposed method has the potential advantages to be applied in practical industry, and future research will explore handling unbalanced real-time datasets in various scenarios. [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]

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]

Semisupervised semantic segmentation of remote sensing images has been proven to be an effective approach to reduce manual annotation costs and leverage available unlabeled data to improve segmentation performance. However, some existing methods that focus on self-training and consistent regularization fail to consider large-scale characteristics of remote sensing images and the importance of incorporating road edge information. In this article, we propose a novel semisupervised edge-aware network (SSEANet) for remote sensing image semantic segmentation by jointly training convolutional neural network and transformer. SSEANet focuses on the consistency loss of multiscale features and uses an attention mechanism to fuse road edge information. Extensive experiments on DeepGlobe, Massachusetts, and AerialKITTI–Bavaria datasets show that the proposed method outperforms state-of-the-art, demonstrating its effectiveness.

Recently, convolutional neural networks (CNNs) have been widely applied in the seismic facies classification. However, even state-of-the-art CNN architectures often encounter classification confusion distinguishing seismic facies at their boundaries. In addition, the annotation is a highly time-consuming task, especially when dealing with 3-D seismic data volumes. While traditional semisupervised methods reduce dependence on annotation, they are susceptible to interference from unreliable pseudolabels. To address these challenges, we propose a semisupervised seismic facies classification method called CONSS, which effectively mitigates classification confusion through contrastive learning. Our proposed method requires only 1% of labeled data, significantly reducing the demand for annotation. To minimize the influence of unreliable pseudolabels, we also introduce a confidence strategy to select positive and negative sample pairs from reliable regions for contrastive learning. Experimental results on the publicly available seismic datasets, the Netherlands F3 and SEAM AI challenge datasets, demonstrate that the proposed method outperforms classic semisupervised methods, including self-training and consistency regularization, achieving exceptional classification performance.

Contrastive learning is a promising approach to unsupervised learning, as it inherits the advantages of well-studied deep models without a dedicated and complex model design. In this paper, based on bidirectional encoder representations from transformers (BERT) and long-short term memory (LSTM) neural networks, we propose self-supervised contrastive learning (SCL) as well as few-shot contrastive learning (FCL) with unsupervised data augmentation (UDA) for text clustering. BERT-SCL outperforms state-of-the-art unsupervised clustering approaches for short texts and for long texts in terms of several clustering evaluation measures. LSTM-SCL also shows good performance for short text clustering. BERT-FCL achieves performance close to supervised learning, and BERT-FCL with UDA further improves the performance for short texts. LSTM-FCL outperforms the supervised model in terms of several clustering evaluation measures. Our experiment results suggest that both SCL and FCL are effective for text clustering.

While various methods have been developed to estimate the age of impact craters, such as the crater size frequency distribution and morphology methods. Accurately and efficiently estimating the ages of lunar craters using traditional techniques is challenging due to their complex morphology and large number. As a result, the accuracy of age estimation algorithms for meteorite craters based on deep learning is restricted by factors such as a scarcity of age-labeled data and the complex morphology of these craters. To address these issues, this article presents an enhanced double-branch network for estimating crater ages via semisupervised learning and multisource lunar data. The algorithm consists of three steps: semisupervised training data augmentation, adaptive two-branch feature extraction, and a two-stage crater age classification process. The effectiveness of the improved approach was validated through ablation experiments, resulting in an overall accuracy of 83.7% on the test set of meteorite craters. This is 5.2% higher than the accuracy achieved by the previous deep learning method.

Semisupervised learning is a forcible method to lessen the cost of annotation for remote sensing semantic segmentation tasks. Recent related research works indicate that consistency training is one of the most effective strategies in semisupervised learning. The core of consistency training is maintaining model outputs consistent under various perturbations. However, the current consistency training-based semisupervised semantic segmentation frameworks lack the analysis of model uncertainty, which increases the generation of semantic ambiguity on remote sensing images. Therefore, we propose the certainty-aware consistency training (CACT) strategy to mitigate the influence of semantic ambiguity caused by model uncertainty. The CACT strategy consists of two novel parts: certainty-aware consistency correction (CACC) and class-balanced-adaptive threshold (CBAT) strategy. The CACC starts with generating a high-quality prediction target, then models the importance of the consistent output target and corrects the output predictions according to the certainty map, increasing the focus on reliable predictions. The CBAT strategy uses a dynamic class-balanced adaptive threshold to filter out unreliable predictions, further reducing the impact of semantic ambiguity. Finally, considerable experimental results on the DLRSD, WHDLD, and Potsdam demonstrate that our framework has an excellent performance on semisupervised remote sensing semantic segmentation scenarios.

Identifying the patterns of defective chips in wafer bin maps (WBMs) in semiconductor manufacturing processes is crucial because different defect patterns correspond to different root causes of process failures. Recently, mixed-type defect patterns (i.e., multiple defect patterns in a single wafer) have become increasingly common owing to the increased complexity of semiconductor manufacturing processes. Previous methods for classifying mixed-type defect patterns in WBMs focused on outputting only the class labels of the defect patterns and not their locations, although location information of the defect patterns is useful for tracking the root causes of failure and improving processes. Moreover, most previous methods used only labeled WBM data, although a larger quantity of unlabeled WBM data are more accessible because of the costly process of label annotation. Therefore, in this paper, we propose a semi-supervised learning method for classifying mixed-type defect patterns and detecting their locations simultaneously using both labeled and unlabeled WBM data. The proposed method extends a recent unsupervised object detection method called Attend-Infer-Repeat in a semi-supervised manner to perform object detection and classification simultaneously. The performance of the proposed method is verified using WBM datasets of different sizes. The results demonstrate the effectiveness of the proposed method for classification and location detection. [ABSTRACT FROM AUTHOR]

COVID-19 is a disease caused by the SARS-CoV-2 virus, with various non-specific symptoms, including cough, fever, shortness of breath, and acute inflammation (hyperinflammation). These symptoms tend to worsen when inflammation is not controlled, and its severe cases can lead to death. Therefore, this study aimed to build a stacked autoencoder deep neural network (SAE-DNN) model for identifying herbal compound candidates that can be used as anti-inflammatory drugs for COVID-19 treatment. The model’s performance was evaluated based on different data representations. The study process involved data collection and preprocessing, modeling, and testing the model on the herbal data to obtain compound candidates. The results showed that the developed SAE-DNN model with compound representation that combined fingerprint and dipeptide composition produced the best performance. Its values are 0.96722, 0.96419, 0.99596, and 0.96567 for accuracy, recall, the area under the receiver operating characteristic, and F1 score, respectively. Furthermore, a total of 33 herbal compounds were identified as potential anti-inflammatory drugs using the SAE-DNN model. [ABSTRACT FROM AUTHOR]

We introduce an active, semisupervised algorithm that utilizes Bayesian experimental design to address the shortage of annotated images required to train and validate Artificial Intelligence (AI) models for lung cancer screening with computed tomography (CT) scans. Our approach incorporates active learning with semisupervised expectation maximization to emulate the human in the loop for additional ground truth labels to train, evaluate, and update the neural network models. Bayesian experimental design is used to intelligently identify which unlabeled samples need ground truth labels to enhance the model's performance. We evaluate the proposed Active Semi-supervised Expectation Maximization for Computer aided diagnosis (CAD) tasks (ASEM-CAD) using three public CT scans datasets: the National Lung Screening Trial (NLST), the Lung Image Database Consortium (LIDC), and Kaggle Data Science Bowl 2017 for lung cancer classification using CT scans. ASEM-CAD can accurately classify suspicious lung nodules and lung cancer cases with an area under the curve (AUC) of 0.94 (Kaggle), 0.95 (NLST), and 0.88 (LIDC) with significantly fewer labeled images compared to a fully supervised model. This study addresses one of the significant challenges in early lung cancer screenings using low-dose computed tomography (LDCT) scans and is a valuable contribution towards the development and validation of deep learning algorithms for lung cancer screening and other diagnostic radiology examinations. [ABSTRACT FROM AUTHOR]