Semi-supervised Learning Framework Research Articles

A novel semi-supervised robust learning framework for dynamic generative latent variable models and its application to industrial virtual metrology

Among a variety of virtual metrology models, dynamic generative latent variable models (DGLVMs) have proven to be an effective tool, owing to outstanding advantages in dealing with complicated data characteristics including temporal correlations, variable colinearities, high dimensionality, and process and measurement uncertainties. Presently, the vast majority of DGLVMs are developed upon the assumption that data are Gaussian-distributed, which is however, non-robust against outlying data and leads the DGLVMs to compromised performance, as the Gaussian distribution is susceptible to outliers that would skew the estimated parameters of DGLVMs. In view of this, this paper proposes a semi-supervised robust learning (SsRL) framework for the DGLVMs by designing an ad-hoc fully robust model structure that handles the outlying data by means of the heavy-tail properties of Student’s t distribution. In addition, the SsRL-DGLVM framework is initiated as a concrete model using probabilistic slow feature analysis (denoted RoSsPSFA) for which a rigorous training algorithm without resorting to any approximations is developed. The performance of the RoSsPSFA is thoroughly evaluated by both numerical and real-world industrial cases, showing the RoSsPSFA’s better immunity against outlying data and higher generalization accuracy.

Extended fiducial inference: toward an automated process of statistical inference

Abstract While fiducial inference was widely considered a big blunder by R.A. Fisher, the goal he initially set—‘inferring the uncertainty of model parameters on the basis of observations’—has been continually pursued by many statisticians. To this end, we develop a new statistical inference method called extended Fiducial inference (EFI). The new method achieves the goal of fiducial inference by leveraging advanced statistical computing techniques while remaining scalable for big data. Extended Fiducial inference involves jointly imputing random errors realized in observations using stochastic gradient Markov chain Monte Carlo and estimating the inverse function using a sparse deep neural network (DNN). The consistency of the sparse DNN estimator ensures that the uncertainty embedded in observations is properly propagated to model parameters through the estimated inverse function, thereby validating downstream statistical inference. Compared to frequentist and Bayesian methods, EFI offers significant advantages in parameter estimation and hypothesis testing. Specifically, EFI provides higher fidelity in parameter estimation, especially when outliers are present in the observations; and eliminates the need for theoretical reference distributions in hypothesis testing, thereby automating the statistical inference process. Extended Fiducial inference also provides an innovative framework for semisupervised learning.

Intelligent Classification and Segmentation of Sandstone Thin Section Image Using a Semi-Supervised Framework and GL-SLIC

This study presents the development and validation of a robust semi-supervised learning framework specifically designed for the automated segmentation and classification of sandstone thin section images from the Yanchang Formation in the Ordos Basin. Traditional geological image analysis methods encounter significant challenges due to the labor-intensive and error-prone nature of manual labeling, compounded by the diversity and complexity of rock thin sections. Our approach addresses these challenges by integrating the GL-SLIC algorithm, which combines Gabor filters and Local Binary Patterns for effective superpixel segmentation, laying the groundwork for advanced component identification. The primary innovation of this research is the semi-supervised learning model that utilizes a limited set of manually labeled samples to generate high-confidence pseudo labels, thereby significantly expanding the training dataset. This methodology effectively tackles the critical challenge of insufficient labeled data in geological image analysis, enhancing the model’s generalization capability from minimal initial input. Our framework improves segmentation accuracy by closely aligning superpixels with the intricate boundaries of mineral grains and pores. Additionally, it achieves substantial improvements in classification accuracy across various rock types, reaching up to 96.3% in testing scenarios. This semi-supervised approach represents a significant advancement in computational geology, providing a scalable and efficient solution for detailed petrographic analysis. It not only enhances the accuracy and efficiency of geological interpretations but also supports broader hydrocarbon exploration efforts.

A comprehensive fault detection and diagnosis method for chemical processes

Deep learning methods based on supervised learning have various applications in chemical process fault detection and diagnosis (FDD). However, these methods as black boxes are unable to provide guidance for process recovery, and the labels and fault samples available in actual chemical processes are insufficient for ideal supervised learning. In this study, eMixDTFCN containing FDD backbone network, semi-supervised learning framework, and explainable method is proposed for the first time to address the above challenges simultaneously. Dense temporal feature convolutional network (DTFCN) combining process variable embedding and dense temporal convolution is taken as the backbone network, where the sparse features from the high-dimensional embedded space are learned efficiently through the dense temporal convolution. MixMatch based on consistent regularization and entropy minimization enables DTFCN with semi-supervised learning and imbalanced learning. SHAP method trains an explainer to provide fault-related variables for process recovery. The high effectiveness and comprehensiveness of eMixDTFCN are validated in two typical chemical processes.

Multi-head co-training: An uncertainty-aware and robust semi-supervised learning framework

Co-training, an advanced form of self-training, allows multiple base models to learn collaboratively, leading to superior performance in semi-supervised learning tasks. However, its widespread adoption is hindered by high computational costs and intricate design choices. To address these challenges, we present Multi-Head Co-Training, a streamlined and efficient framework that consolidates individual models into a multi-head structure, adding minimal extra parameters. Each classification head in this unified model collaborates with others via a ”Weak and Strong Augmentation” strategy, with diversity organically introduced through robust data augmentation. Consequently, our approach implicitly promotes diversity while incurring only a minor increase in computational overhead, making co-training more accessible. We validate the effectiveness of Multi-Head Co-Training through an empirical study on standard semi-supervised learning benchmarks. For example, our method achieves up to a 3.1% accuracy improvement on the semi-supervised CIFAR dataset compared to recent methods.Recognizing the necessity for more practical performance metrics beyond accuracy, we assess our framework from three additional perspectives: robust generalization, uncertainty, and computational efficiency. To evaluate robust generalization, we expand the conventional SSL experimental setting to a more comprehensive open-set semi-supervised learning scenario. For uncertainty assessment, we conduct experiments on model calibration and selective classification benchmarks. For example, our method achieves up to a 4.3% accuracy improvement on the open-set semi-supervised CIFAR dataset. Our extensive experiments confirm that our proposed framework better captures prediction confidence and uncertainty, rendering it more suitable for SSL deployment in open environments. The code is available at https://github.com/chenmc1996/Multi-Head-Co-Training.

PanIN and CAF transitions in pancreatic carcinogenesis revealed with spatial data integration

This study introduces a new imaging, spatial transcriptomics (ST), and single-cell RNA-sequencing integration pipeline to characterize neoplastic cell state transitions during tumorigenesis. We applied a semi-supervised analysis pipeline to examine premalignant pancreatic intraepithelial neoplasias (PanINs) that can develop into pancreatic ductal adenocarcinoma (PDAC). Their strict diagnosis on formalin-fixed and paraffin-embedded (FFPE) samples limited the single-cell characterization of human PanINs within their microenvironment. We leverage whole transcriptome FFPE ST to enable the study of a rare cohort of matched low-grade (LG) and high-grade (HG) PanIN lesions to track progression and map cellular phenotypes relative to single-cell PDAC datasets. We demonstrate that cancer-associated fibroblasts (CAFs), including antigen-presenting CAFs, are located close to PanINs. We further observed a transition from CAF-related inflammatory signaling to cellular proliferation during PanIN progression. We validate these findings with single-cell high-dimensional imaging proteomics and transcriptomics technologies. Altogether, our semi-supervised learning framework for spatial multi-omics has broad applicability across cancer types to decipher the spatiotemporal dynamics of carcinogenesis.

On-Device Semi-Supervised Activity Detection: A New Privacy-Aware Personalized Health Monitoring Approach.

This paper presents an on-device semi-supervised human activity detection system that can learn and predict human activity patterns in real time. The clinical objective is to monitor and detect the unhealthy sedentary lifestyle of a user. The proposed semi-supervised learning (SSL) framework uses sparsely labelled user activity events acquired from Inertial Measurement Unit sensors installed as wearable devices. The proposed cluster-based learning model in this approach is trained with data from the same target user, thus preserving data privacy while providing personalized activity detection services. Two different cluster labelling strategies, namely, population-based and distance-based strategies, are employed to achieve the desired classification performance. The proposed system is shown to be highly accurate and computationally efficient for different algorithmic parameters, which is relevant in the context of limited computing resources on typical wearable devices. Extensive experimentation and simulation study have been conducted on multi-user human activity data from the public domain in order to analyze the trade-off between classification accuracy and computation complexity of the proposed learning paradigm with different algorithmic hyper-parameters. With 4.17 h of training time for 8000 activity episodes, the proposed SSL approach consumes at most 20 KB of CPU memory space, while providing a maximum accuracy of 90% and 100% classification rates.

Wind turbine blade defect detection with a semi-supervised deep learning framework

To increase the economic efficiency of utilizing wind turbines, an enhanced object detection model based on the small version of the fifth generation of the You Look Only Once algorithm (YOLOv5s) is proposed in this paper, which can effectively detect cracks on the surface of wind turbine blades using images captured by unmanned aerial vehicles. To improve the extraction of light-colored and low-definition images, Omni-Dimensional Dynamic Convolution (ODConv) and Dynamic Head Module components (DyHead) are introduced in the proposed method, while a lightweight Group Shuffle convolution (GSConv) module is utilized to accelerate the model inference speed without sacrificing detection performance. Furthermore, a semi-supervised learning strategy is developed to reduce human labors in annotating images. Extensive experiments demonstrate that the proposed model outperforms the original YOLOv5s in terms of both detection accuracy and inference speed. Besides, the proposed model has good performance against state-of-the-art methods. Furthermore, the experiments validate the efficacy of the proposed semi-supervised learning strategy.

Discriminative atoms embedding relation dual network for classification of choroidal neovascularization in OCT images

Choroidal neovascularization (CNV) is an eye disease that can cause vision loss. Automatic CNV classification in OCT images is crucial in the treatment of CNV. However, two problems arise for CNV classification in OCT images. The subtle visual differences between different CNV types render classification difficult. Additionally, it is difficult to obtain sufficient labeled data, which results in performance degradation. In order to solve these two problems, a discriminative atom-embedding relation dual network is proposed in this paper. Considering that semi-supervised learning (SSL) is an effective machine learning framework to make full use of limited labeled data and a large amount of unlabeled data, the proposed network is developed within an SSL framework. To capture the visual differences, novel discriminative atoms are first introduced to mine discriminative information between different CNV types. Subsequently, a relation module is incorporated to embed the learned discriminative atom information into the features. This makes the learned features capable of distinguishing between different CNV types. Moreover, a novel relation consistency loss is proposed to further improve the robustness of the learned features. Experimental results on private and public datasets demonstrate the effectiveness of the proposed method.

BearingFM: Towards a foundation model for bearing fault diagnosis by domain knowledge and contrastive learning

Monitoring bearing failures in production equipment can effectively prevent finished product quality issues and unplanned factory downtime, thereby reducing supply chain uncertainties and risk. Therefore, monitoring bearing failures in production equipment is important for improving supply chain sustainability. Due to the generalization limitations of neural network models, specific models must be trained for specific tasks. However, in real industrial scenarios, there is a severe lack of labeled samples, making it difficult to deploy fault diagnosis models across massive amounts of equipment in workshops. In order to solve the above issue, this paper proposes a cloud-edge-end collaborative semi-supervised learning framework, which provides multi-level computing power and data support for building a foundation model. A data augmentation method based on the bearing fault mechanism is proposed, which effectively preserves the inherent essential characteristics in vibration signals by normalizing frequency and adding noise in specific frequency bands. A novel contrastive learning model is designed, which narrows the distances between positive samples and widens the distances between negative samples in the high-dimensional space through cross comparisons in the time dimension and knowledge dimension, thereby extracting the most essential characteristics from the unlabeled signals. Multiple sets of experiments conducted on four datasets demonstrate that the proposed approach achieves an approximately 98% fault classification accuracy with only 1.2% labeled samples.

Leveraging a self-adaptive mean teacher model for semi-supervised multi-exposure image fusion

Deep learning-based methods have recently shown remarkable advancements in multi-exposure image fusion (MEF), demonstrating significant achievements in improving the fusion quality. Despite their success, the majority of reference images in MEF are artificially generated, inevitably introducing a portion of low-quality ones. Existing methods either utilize these mixed-quality reference images for supervised learning or heavily depend on source images for unsupervised learning, making the fusion results challenging to accurately reflect real-world illumination conditions. To overcome the impact of unreliable factors in references, we propose a self-adaptive mean teacher-based semi-supervised learning framework tailored for MEF, termed SAMT-MEF. Its self-adaptiveness is reflected from two perspectives. Firstly, we establish a self-adaptive set to retain the best-ever outputs from the teacher as pseudo labels, employing a well-crafted hybrid metric for its updates. Secondly, we employ contrastive learning to assist the self-adaptive set further in alleviating overfitting to inferior pseudo labels. Our proposed method, backed by abundant empirical evidence, outperforms state-of-the-art methods quantitatively and qualitatively on both reference and non-reference datasets. Furthermore, in some scenarios, the fusion results surpass the reference images, showcasing superior performance in practical applications. Source code are publicly available at https://github.com/hqj9994ever/SAMT-MEF.

Semi‐supervised learning framework with shape encoding for neonatal ventricular segmentation from 3D ultrasound

AbstractBackgroundThree‐dimensional (3D) ultrasound (US) imaging has shown promise in non‐invasive monitoring of changes in the lateral brain ventricles of neonates suffering from intraventricular hemorrhaging. Due to the poorly defined anatomical boundaries and low signal‐to‐noise ratio, fully supervised methods for segmentation of the lateral ventricles in 3D US images require a large dataset of annotated images by trained physicians, which is tedious, time‐consuming, and expensive. Training fully supervised segmentation methods on a small dataset may lead to overfitting and hence reduce its generalizability. Semi‐supervised learning (SSL) methods for 3D US segmentation may be able to address these challenges but most existing SSL methods have been developed for magnetic resonance or computed tomography (CT) images.PurposeTo develop a fast, lightweight, and accurate SSL method, specifically for 3D US images, that will use unlabeled data towards improving segmentation performance.MethodsWe propose an SSL framework that leverages the shape‐encoding ability of an autoencoder network to enforce complex shape and size constraints on a 3D U‐Net segmentation model. The autoencoder created pseudo‐labels, based on the 3D U‐Net predicted segmentations, that enforces shape constraints. An adversarial discriminator network then determined whether images came from the labeled or unlabeled data distributions. We used 887 3D US images, of which 87 had manually annotated labels and 800 images were unlabeled. Training/validation/testing sets of 25/12/50, 25/12/25 and 50/12/25 images were used for model experimentation. The Dice similarity coefficient (DSC), mean absolute surface distance (MAD), and absolute volumetric difference (VD) were used as metrics for comparing to other benchmarks. The baseline benchmark was the fully supervised vanilla 3D U‐Net while dual task consistency, shape‐aware semi‐supervised network, correlation‐aware mutual learning, and 3D U‐Net Ensemble models were used as state‐of‐the‐art benchmarks with DSC, MAD, and VD as comparison metrics. The Wilcoxon signed‐rank test was used to test statistical significance between algorithms for DSC and VD with the threshold being p < 0.05 and corrected to p < 0.01 using the Bonferroni correction. The random‐access memory (RAM) trace and number of trainable parameters were used to compare the computing efficiency between models.ResultsRelative to the baseline 3D U‐Net model, our shape‐encoding SSL method reported a mean DSC improvement of 6.5%, 7.7%, and 4.1% with a 95% confidence interval of 4.2%, 5.7%, and 2.1% using image data splits of 25/12/50, 25/12/25, and 50/12/25, respectively. Our method only used a 1GB increase in RAM compared to the baseline 3D U‐Net and required less than half the RAM and trainable parameters compared to the 3D U‐Net ensemble method.ConclusionsBased on our extensive literature survey, this is one of the first reported works to propose an SSL method designed for segmenting organs in 3D US images and specifically one that incorporates unlabeled data for segmenting neonatal cerebral lateral ventricles. When compared to the state‐of‐the‐art SSL and fully supervised learning methods, our method yielded the highest DSC and lowest VD while being computationally efficient.

A comprehensive evaluation method for frailty based on semi-supervised learning and transfer-learning

Frailty evaluation is of great significance for the specific population, which can speed up the treatment process and reduce the adverse effects after treatment. In this article, in order to make up for the shortcomings of the traditional evaluation methods, a comprehensive intelligent evaluation method that integrates physiological data from multiple perspectives, such as mobility, muscle level, and body composition data has been proposed. The multi-source physiological data of 174 participants has been collected, and a series of features related to frailty have been extracted. A recurrent stepping semi-supervised learning (RSSSL) framework is proposed to transmit the knowledge from the annotated data to the unannotated gait sequence utilizing various networks with varying architectures. Moreover, in order to enhance the frailty evaluation’s performance, an approach based on subspace transfer learning integrating with heterogeneous features (STL-IHF) is proposed. The mean segmentation accuracy on the unannotated gait sequence reaches over 97.10% across recruited subjects with distinct gait patterns, which is significantly superior when compared with classical deep learning methods. And the recognition accuracy of the STL-IHF method for frailty evaluation reaches over 92.9%. The overall experimental results demonstrate that the proposed methods are effective in frailty evaluation tasks in clinics.

A Semi-Supervised Learning Framework Combining CNN and Multiscale Transformer for Traffic Sign Detection and Recognition

A Semi-Supervised Learning Framework Combining CNN and Multiscale Transformer for Traffic Sign Detection and Recognition

Triplet Adaptation Framework for Robust Semi-Supervised Learning.

Semi-supervised learning (SSL) suffers from severe performance degradation when labeled and unlabeled data come from inconsistent and imbalanced distribution. Nonetheless, there is a lack of theoretical guidance regarding a remedy for this issue. To bridge the gap between theoretical insights and practical solutions, we embark to an analysis of generalization bound of classic SSL algorithms. This analysis reveals that distribution inconsistency between unlabeled and labeled data can cause a significant generalization error bound. Motivated by this theoretical insight, we present a Triplet Adaptation Framework (TAF) to reduce the distribution divergence and improve the generalization of SSL models. TAF comprises three adapters: Balanced Residual Adapter, aiming to map the class distribution of labeled and unlabeled data to a uniform distribution for reducing class distribution divergence; Representation Adapter, aiming to map the representation distribution of unlabeled data to labeled one for reducing representation distribution divergence; and Pseudo-Label Adapter, aiming to align the predicted pseudo-labels with the class distribution of unlabeled data, thereby preventing erroneous pseudo-labels from exacerbating representation divergence. These three adapters collaborate synergistically to reduce the generalization bound, ultimately achieving a more robust and generalizable SSL model. Extensive experiments across various robust SSL scenarios validate the efficacy of our method.

A heterogeneous graph-based semi-supervised learning framework for access control decision-making

For modern information systems, robust access control mechanisms are vital in safeguarding data integrity and ensuring the entire system’s security. This paper proposes a novel semi-supervised learning framework that leverages heterogeneous graph neural network-based embedding to encapsulate both the intricate relationships within the organizational structure and interactions between users and resources. Unlike existing methods focusing solely on individual user and resource attributes, our approach embeds organizational and operational interrelationships into the hidden layer node embeddings. These embeddings are learned from a self-supervised link prediction task based on a constructed access control heterogeneous graph via a heterogeneous graph neural network. Subsequently, the learned node embeddings, along with the original node features, serve as inputs for a supervised access control decision-making task, facilitating the construction of a machine-learning access control model. Experimental results on the open-sourced Amazon access control dataset demonstrate that our proposed framework outperforms models using original or manually extracted graph-based features from previous works. The prepossessed data and codes are available on GitHub,facilitating reproducibility and further research endeavors.

DDformer: Dimension decomposition transformer with semi-supervised learning for underwater image enhancement

Vision-guided Autonomous Underwater Vehicles (AUVs) have gradually become significant tools for human exploration of the ocean. However, distorted images severely limit the visual ability, making it difficult to meet the needs of complex underwater environment perception. Fortunately, recent advancements in deep learning have led to rapid developments in underwater image enhancement. The emergence of the Transformer architecture has further enhanced the capabilities of deep learning. However, the direct application of Transformer to underwater image enhancement presents challenges in computing pixel-level global information and extracting local features. In this paper, we present a novel approach that merges dimension decomposition Transformer with semi-supervised learning for underwater image enhancement. To begin, dimension decomposition attention is proposed, which enables Transformer to compute global dependencies directly at the original scale and correct color distortions effectively. Concurrently, we employ convolutional neural networks to compensate for Transformer's limitations in extracting local features, thereby enriching details and textures. Subsequently, a multi-stage Transformer strategy is introduced to divide the network into high- and low-resolution stages for multi-scale global information extraction. It helps correct color distortions while enhancing the network's focus on regions with severe degradation. Moreover, we design a semi-supervised learning framework to reduce the reliance on paired datasets and construct a corresponding multi-scale fusion discriminator to enhance the sensitivity to input data. Experimental results demonstrate that our method outperforms state-of-the-art approaches, showcasing excellent learning and generalization capabilities on subjective perception and overall evaluation metrics. Furthermore, outstanding results highlight the significant improvements it brings to downstream visual engineering applications. The code of the proposed DDformer is available at https://github.com/ZhiGao-hfuu/DDformer.

An enhanced semi-supervised learning method with self-supervised and adaptive threshold for fault detection and classification in urban power grids

With the rapid development of urban power grids and the large-scale integration of renewable energy, traditional power grid fault diagnosis techniques struggle to address the complexities of diagnosing faults in intricate power grid systems. Although artificial intelligence technologies offer new solutions for power grid fault diagnosis, the difficulty in acquiring labeled grid data limits the development of AI technologies in this area. In response to these challenges, this study proposes a semi-supervised learning framework with self-supervised and adaptive threshold (SAT-SSL) for fault detection and classification in power grids. Compared to other methods, our method reduces the dependence on labeling data while maintaining high recognition accuracy. First, we utilize frequency domain analysis on power grid data to filter abnormal events, then classify and label these events based on visual features, to creating a power grid dataset. Subsequently, we employ the Yule–Walker algorithm extract features from the power grid data. Then we construct a semi-supervised learning framework, incorporating self-supervised loss and dynamic threshold to enhance information extraction capabilities and adaptability across different scenarios of the model. Finally, the power grid dataset along with two benchmark datasets are used to validate the model’s functionality. The results indicate that our model achieves a low error rate across various scenarios and different amounts of labels. In power grid dataset, When retaining just 5% of the labels, the error rate is only 6.15%, which proves that this method can achieve accurate grid fault detection and classification with a limited amount of labeled data.

AutoDLAR: A Semi-supervised Cross-modal Contact-free Human Activity Recognition System

WiFi-based human activity recognition (HAR) plays an essential role in various applications such as security surveillance, health monitoring, and smart home. Existing HAR methods, though yielding promising performance in indoor scenarios, highly depend on a massive labeled dataset for training which is extremely difficult to acquire in practical applications. In this paper, we present an automatic data labeling and HAR system, termed AutoDLAR. Taking a semi-supervised cross-modal learning framework with a hybrid loss function as the core, AutoDLAR transfers rich visual information to automatically label WiFi signals for WiFi-based HAR. Specifically, we devise a lightweight and multi-view WiFi sensing model with a parallel feature embedding method to accurately identify activities and accelerate recognition speed. Then, we exploit the video data to fine-tune a well-established visual HAR model, generating effective pseudo-labels for guiding the WiFi model’s training. We also build a synchronized Video-WiFi dataset with seven types of human activities under different scenarios to enable training and validating the semi-supervised HAR system. Extensive experiments on our collected activity dataset and the emotion recognition benchmark demonstrate that AutoDLAR attains an average accuracy of over 95.89% without manual labeling and only spends the inference time of 3.35 ms, outperforming the state-of-the-art (SOTA) methods.

Phenome-wide identification of therapeutic genetic targets, leveraging knowledge graphs, graph neural networks, and UK Biobank data.

The ongoing expansion of human genomic datasets propels therapeutic target identification; however, extracting gene-disease associations from gene annotations remains challenging. Here, we introduce Mantis-ML 2.0, a framework integrating AstraZeneca's Biological Insights Knowledge Graph and numerous tabular datasets, to assess gene-disease probabilities throughout the phenome. We use graph neural networks, capturing the graph's holistic structure, and train them on hundreds of balanced datasets via a robust semi-supervised learning framework to provide gene-disease probabilities across the human exome. Mantis-ML 2.0 incorporates natural language processing to automate disease-relevant feature selection for thousands of diseases. The enhanced models demonstrate a 6.9% average classification power boost, achieving a median receiver operating characteristic (ROC) area under curve (AUC) score of 0.90 across 5220 diseases from Human Phenotype Ontology, OpenTargets, and Genomics England. Notably, Mantis-ML 2.0 prioritizes associations from an independent UK Biobank phenome-wide association study (PheWAS), providing a stronger form of triaging and mitigating against underpowered PheWAS associations. Results are exposed through an interactive web resource.