Search Results (33)

Visually rich document understanding (VrDU) models rely on tightly coupled textual, layout, and visual representations. In real-world deployments, these models must continuously adapt to new document domains over time. However, naïve sequential fine-tuning leads to severe catastrophic forgetting due to shared parameters and strong cross-task interference. Existing continual learning approaches either constrain parameter updates, preserve output distributions, or uniformly suppress gradient directions associated with previous tasks. While effective in limited settings, these strategies fail to balance stability and plasticity in large multimodal transformers. We propose AMD-Proj, an adaptive memory-driven selective gradient projection framework for continual learning in document understanding. It models task knowledge using specific gradient subspaces and adaptively modulates incoming gradients based on their alignment with this memory, selectively blocking interfering directions while reinforcing reusable ones. An efficient truncated SVD mechanism with online subspace merging ensures bounded memory usage and scalability to large transformer-based architectures. We evaluate AMD-Proj on four VrDU benchmarks (FUNSD, SROIE, CORD, and BuDDIE) under a task-incremental learning setting using LayoutLMv2 and LayoutLMv3 backbones. Results show that AMD-Proj reduces catastrophic forgetting and improves F1-based stability over EWC, GPM, LwF, OWM, CUBER, TRGP and parameter-efficient fine-tuning methods. Extensive mechanistic analyses, including gradient spectrum decomposition and layer-wise reuse versus block dynamics, provide insight into how selective gradient projection controls optimization geometry during continual adaptation. These findings establish selective gradient projection as a principled and interpretable approach for continual learning in visually rich document understanding. Full article

Ballast compaction is a critical factor influencing ballast bed condition and the operational safety of heavy-haul railways. However, existing quantitative evaluation methods often suffer from overly idealized simulation models and limitations in signal processing and assessment frameworks. To address these issues, this study proposes a quantitative analysis approach for ballast compaction by integrating non-uniform medium simulation modeling, wavelet–Singular Value Decomposition (SVD) joint denoising, frequency–wavenumber (F-K) migration imaging, and wavelet packet decomposition (WPD)-based feature extraction. Forward simulations were conducted based on the constructed model, and the proposed methodology was validated using 1.5 GHz (gigahertz, 1 GHz = ${10}^9$ Hz) ground penetrating radar (GPR) data acquired from compaction experiments. The results demonstrate that wavelet–SVD joint denoising effectively suppresses deep coherent noise caused by strong reflections from sleepers, significantly enhancing the identification of deep effective signals and ensuring accurate localization and feature extraction of compaction zones. The Geometric Mean of WPD High/Low-Frequency Energy Ratio (GMHLFER) exhibits a strong positive correlation with the degree of compaction. In simulations, as the proportion of compacted material increased from $9\%$ to $21\%$, the GMHLFER rose from 21.555 to 26.581. In field tests, the value increased from 22.012 to 26.012 as compaction severity progressed from slight to severe, demonstrating stable responses across full-gradient compaction conditions and indicating high robustness and sensitivity. The proposed method provides an effective approach for quantitative characterization of ballast compaction in heavy-haul railways, and offers a promising technical pathway for intelligent inspection and condition assessment of railway ballast beds. Full article

Background/Objectives: Acute health events in institutionalized older adults often lead to avoidable hospital referrals, requiring rapid, accurate remote decision-making. Telemedicine has become a key tool to improve assessment and care continuity in nursing homes. This study aimed to evaluate outcomes associated with telemedicine-supported management of acute events in residential care facilities for older adults and to develop a deep learning model to classify episodes and predict hospital referrals. Methods: A quasi-experimental study analyzed 5202 acute events managed via a 24/7 telemedicine system in Vitalia nursing homes (January–October 2024). The dataset included demographics, comorbidities, vital signs, event characteristics, and outcomes. Data preprocessing involved imputation, normalization, encoding, and dimensionality reduction via Truncated SVD (200 components). Given the imbalance in referral outcomes (~10%), several resampling techniques (SMOTE, SMOTEENN, SMOTETomek) were applied. A deep feedforward neural network (256–128–64 units with Batch Normalization, LeakyReLU, Dropout, AdamW) was trained using stratified splits (70/10/20) and optimized via cross-validation. Results: Telemedicine enabled the resolution of approximately 90% of acute events within the residential setting, reducing reliance on emergency services. The deep learning model outperformed traditional algorithms, achieving its best performance with SMOTEENN preprocessing (AUC = 0.91, accuracy = 0.88). The proposed model achieved higher overall performance than baseline classifiers, providing a more balanced precision–specificity trade-off for hospital referral prediction, with an F1-score of 0.63. Conclusions: Telemedicine-enabled acute care, strengthened by a robust deep learning classifier, offers a reliable strategy to enhance triage accuracy, reduce unnecessary transfers, and optimize clinical decision-making in nursing homes. These findings support the integration of AI-assisted telemedicine systems into long-term care workflows. Full article

Structural valve deterioration (SVD) remains a key limitation in bioprosthetic heart valve (BHV) usage influenced by patient age. A deeper understanding of SVD pathogenesis, particularly of the immune-mediated processes altering current BHV materials, is therefore critical. To this end, commercially available BHV tissues of bovine, porcine, and equine origin were investigated following subcutaneous implantation into α1,3-galactosyltransferase-knockout (Gal KO) mice. We compared the immune responses between adult and aged animals via histological assessments of explants and measurement of serum anti-galactose α1,3-galactose (Gal) and anti-non-Gal antibodies at 2 months post-implantation. In contrast to adult mice, old Gal KO mice did not show increased levels of serum anti-Gal or -non-Gal antibodies after receiving specific BHV tissue (i.e., Freedom-Solo). Instead, a significant decrease in serum anti-Gal IgM was found in old recipients of Freedom-Solo. Furthermore, the overall cellular immune response was attenuated in explants from old mice compared with adults (i.e., ATS 3f and Crown). Nevertheless, the Freedom-Solo (bovine) and the Hancock-II (porcine) tissues still elicited strong cellular immune infiltration in the old cohorts. Therefore, the Gal KO mouse model offers a valuable platform to investigate age-related differences regarding cellular and humoral immune responses to various BHV tissues, contributing to our understanding of SVD. Full article

Background: Cancer poses a significant health risk to humans, with prostate cancer (PCa) being the second most common and deadly form among men, following lung cancer. Each year, it affects over a million individuals and presents substantial diagnostic challenges due to variations in tissue appearance and imaging quality. In recent decades, various techniques utilizing Magnetic Resonance Imaging (MRI) have been developed for identifying and classifying PCa. Accurate classification in MRI typically requires the integration of complementary feature types, such as deep semantic representations from Convolutional Neural Networks (CNNs) and handcrafted descriptors like Histogram of Oriented Gradients (HOG). Therefore, a more robust and discriminative feature integration strategy is crucial for enhancing computer-aided diagnosis performance. Objectives: This study aims to develop a multi-stage hybrid learning model that combines deep and handcrafted features, investigates various feature reduction and classification techniques, and improves diagnostic accuracy for prostate cancer using magnetic resonance imaging. Methods: The proposed framework integrates deep features extracted from convolutional architectures with handcrafted texture descriptors to capture both semantic and structural information. Multiple dimensionality reduction methods, including singular value decomposition (SVD), were evaluated to optimize the fused feature space. Several machine learning (ML) classifiers were benchmarked to identify the most effective diagnostic configuration. The overall framework was validated using k-fold cross-validation to ensure reliability and minimize evaluation bias. Results: Experimental results on the Transverse Plane Prostate (TPP) dataset for binary classification tasks showed that the hybrid model significantly outperformed individual deep or handcrafted approaches, achieving superior accuracy of 99.74%, specificity of 99.87%, precision of 99.87%, sensitivity of 99.61%, and F1-score of 99.74%. Conclusions: By combining complementary feature extraction, dimensionality reduction, and optimized classification, the proposed model offers a reliable and generalizable solution for prostate cancer diagnosis and demonstrates strong potential for integration into intelligent clinical decision-support systems. Full article

Voice biometric systems play a critical role in numerous security applications, including electronic device authentication, banking transaction verification, and confidential communications. Despite their widespread utility, these systems are increasingly targeted by sophisticated spoofing attacks that leverage advanced artificial intelligence techniques to generate realistic synthetic speech. Addressing the vulnerabilities inherent to voice-based authentication systems has thus become both urgent and essential. This study proposes a novel experimental analysis that extensively explores various dimensionality reduction strategies in conjunction with supervised machine learning models to effectively identify spoofed voice signals. Our framework involves extracting multicepstral features followed by the application of diverse dimensionality reduction methods, such as Principal Component Analysis (PCA), Truncated Singular Value Decomposition (SVD), statistical feature selection (ANOVA F-value, Mutual Information), Recursive Feature Elimination (RFE), regularization-based LASSO selection, Random Forest feature importance, and Permutation Importance techniques. Empirical evaluation using the ASVSpoof 2017 v2.0 dataset measures the classification performance with the Equal Error Rate (EER) metric, achieving values of approximately 10%. Our comparative analysis demonstrates significant performance gains when dimensionality reduction methods are applied, underscoring their value in enhancing the security and effectiveness of voice biometric verification systems against emerging spoofing threats. Full article

Robust indoor robot navigation typically demands either costly sensors or extensive training data. We propose a cost-effective RGB-D navigation pipeline that couples feature-based relative pose estimation with a lightweight multi-layer-perceptron (MLP) policy. RGB-D keyframes extracted from human-driven traversals form nodes of a topological map; edges are added when visual similarity and geometric–kinematic constraints are jointly satisfied. During autonomy, LightGlue features and SVD give six-DoF relative pose to the active keyframe, and the MLP predicts one of four discrete actions. Low visual similarity or detected obstacles trigger graph editing and Dijkstra replanning in real time. Across eight tasks in four Habitat-Sim environments, the agent covered 190.44 m, replanning when required, and consistently stopped within 0.1 m of the goal while running on commodity hardware. An information-theoretic analysis over the Multi-Illumination dataset shows that LightGlue maximizes per-second information gain under lighting changes, motivating its selection. The modular design attains reliable navigation without metric SLAM or large-scale learning, and seamlessly accommodates future perception or policy upgrades. Full article

The wear condition of a piston pin is a main factor in determining the operational continuity and life cycle of a diesel engine; identifying its vibration feature is of paramount importance in carrying out necessary maintenance in the early wear stage. As the dynamic vibration features are susceptible to environmental disturbance during operation, an effective signal processing method is necessary to improve the accuracy and fineness of the extracted features, which is essential to build a reliable and precise binary classifier model to identify piston pin wear based on the features. Aiming at the feature extraction requirements of anti-noise, accuracy and effectiveness, this paper proposes a piston pin wear feature extraction algorithm based on dynamic principal component analysis (DPCA) combined with variational mode decomposition (VMD) and singular value decomposition (SVD). An orthogonal sensor layout is applied to collect the vibration signal under normal and worn piston pin conditions, which proved effective in reducing environmental vibration disturbance. DPCA is utilized to extract dynamical vibration features by introducing time lag. Then, the dynamic principal component matrix is further decomposed by VMD to obtain intrinsic mode functions (IMFs) as finer features and is finally decomposed by SVD to compress the features, thus improving the classification efficiency based on the features. To validate the significance of the features extracted by the proposed method, a support vector machine (SVM) is employed to model binary classifiers to evaluate the classification performance trained by different features. A modeling dataset containing 80 samples (40 normal samples and 40 worn samples) is employed, and five-round cross-validation is adopted. For each round, two binary classifier models are trained by features extracted by the proposed method and the empirical mode decomposition (EMD)–auto regressive (AR) spectrum method, fast Fourier transform (FFT) and continuous wavelet transform (CWT), respectively; the classification precision, recall ratio, accuracy and F1 ratio are obtained on the testing set by contrasting the overall performances of the five-round cross-validation, and the proposed method is proved to be more effective in noise reduction and significant feature extraction, which is able to improve the accuracy and efficiency of binary classification for piston pin wear identification. Full article

User preferences are important for electric vehicle charging station (EVCS) recommendations, but they have not been deeply analyzed. Therefore, in this study, user charging preferences are identified and applied to EVCS recommendations using a hybrid model that integrates LightGBM and singular value decomposition (SVD). In the model, LightGBM is used to predict user ratings according to users’ comments regarding charging orders, and the feature importance reported by each user is output. Then, a co-occurrence matrix between users and charging stations (EVCSs) is constructed and decomposed using SVD. Based on the decomposed results, the final evaluated scores of each user for EVCSs can be calculated. Upon ranking the EVCSs according to the scores, the EVCS recommendation results are obtained, taking into account the users’ charging preferences. The sample data consist of 28,306 orders from 508 users at 241 charging stations in Linyi, Shandong, China. The experimental results show that the proposed hybrid model outperforms the benchmark models in terms of precision, recall, and F1 score, and its F1 score can be increased by 96% compared with that of the traditional item-based collaborative filtering method with charging counts for EVCS recommendations. Full article

Large-scale symmetric and asymmetric matrices have emerged in predicting the relationship between genes and diseases. The emergence of large-scale matrices increases the computational complexity of the problem. Therefore, using low-rank matrices instead of original symmetric and asymmetric matrices can greatly reduce computational complexity. In this paper, we propose an approximation conjugate gradient method for solving the low-rank matrix recovery problem, i.e., the low-rank matrix is obtained to replace the original symmetric and asymmetric matrices such that the approximation error is the smallest. The conjugate gradient search direction is given through matrix addition and matrix multiplication. The new conjugate gradient update parameter is given by the F-norm of matrix and the trace inner product of matrices. The conjugate gradient generated by the algorithm avoids SVD decomposition. The backtracking linear search is used so that the approximation conjugate gradient direction is computed only once, which ensures that the objective function decreases monotonically. The global convergence and local superlinear convergence of the algorithm are given. The numerical results are reported and show the effectiveness of the algorithm. Full article

This paper presents a novel approach to mind-wandering prediction in the context of webcam-based online learning. We implemented a Singular Value Decomposition (SVD)-based 1D temporal eye-signal extraction method, which relies solely on eye landmark detection and eliminates the need for gaze tracking or specialized hardware, then extract suitable features from the signals to train the prediction model. Our thorough experimental framework facilitates the evaluation of our approach alongside baseline models, particularly in the analysis of temporal eye signals and the prediction of attentional states. Notably, our SVD-based signal captures both subtle and major eye movements, including changes in the eye boundary and pupil, surpassing the limited capabilities of eye aspect ratio (EAR)-based signals. Our proposed model exhibits a 2% improvement in the overall Area Under the Receiver Operating Characteristics curve (AUROC) metric and 7% in the F1-score metric for ‘not-focus’ prediction, compared to the combination of EAR-based and computationally intensive gaze-based models used in the baseline study These contributions have potential implications for enhancing the field of attentional state prediction in online learning, offering a practical and effective solution to benefit educational experiences. Full article

Early-stage Alzheimer’s disease (AD) and frontotemporal dementia (FTD) share similar symptoms, complicating their diagnosis and the development of specific treatment strategies. Our study evaluated multiple feature extraction techniques for identifying AD and FTD biomarkers from electroencephalographic (EEG) signals. We developed an optimised machine learning architecture that integrates sliding windowing, feature extraction, and supervised learning to distinguish between AD and FTD patients, as well as from healthy controls (HCs). Our model, with a 90% overlap for sliding windowing, SVD entropy for feature extraction, and K-Nearest Neighbors (KNN) for supervised learning, achieved a mean F1-score and accuracy of 93% and 91%, 92.5% and 93%, and 91.5% and 91% for discriminating AD and HC, FTD and HC, and AD and FTD, respectively. The feature importance array, an explainable AI feature, highlighted the brain lobes that contributed to identifying and distinguishing AD and FTD biomarkers. This research introduces a novel framework for detecting and discriminating AD and FTD using EEG signals, addressing the need for accurate early-stage diagnostics. Furthermore, a comparative evaluation of sliding windowing, multiple feature extraction, and machine learning methods on AD/FTD detection and discrimination is documented. Full article

Ground-penetrating radar (GPR) is a classic geophysical exploration method that utilizes the emission and reception of electromagnetic waves to non-destructively detect target objects in the target medium. It has been widely applied in various fields such as pipeline detection, cavity detection, and rebar detection. However, GPR systems are susceptible to environmental clutter interference, which poses challenges for data interpretation and subsequent processing. In this paper, the separability of clutter and target signal in the frequency-wavenumber (F-K) domain is validated through modeling, leading to the proposal of a comprehensive clutter removal method based on the F-K domain for complex scenarios. The direct coupling wave is initially eliminated by applying a peak matching mean subtraction filter, which avoids the artifacts. Subsequently, the F-K domain transformation is performed and surface clutter undulations are effectively removed using a method based on singular value decomposition and k-means clustering. Finally, an angle filter with Gaussian tapering at the edges is designed based on physical models to efficiently eliminate linear interference without undesired ringing interference. The commonly used clutter removal algorithms, including mean subtraction (MS), singular value decomposition (SVD), robust principal component analysis (RPCA), and traditional F-K filtering methods, are compared with the proposed algorithm on both the numerical simulated data and actual GPR data. The results from visual and quantitative analysis confirm that our proposed method is more effective than current commonly used clutter suppression algorithms. We have successfully enhanced the Signal-to-Clutter Ratio (SCR) of the GPR data, resulting in an Improvement Factor (IF) of 30.63 dB, 23.59 dB, and 30.60 dB for simulated data, experimental data, and TU1208 public data, respectively. The detection capability of buried targets is enhanced, thereby establishing a solid foundation for subsequent data interpretation and target identification. Full article

In the era of big data, effective recommendation systems are essential for providing users with personalized content and reducing search time on online platforms. Traditional collaborative filtering (CF) methods face challenges like data sparsity and the new-user or cold-start issue, primarily due to their reliance on limited user–item interactions. This paper proposes an innovative movie recommendation system that integrates deep reinforcement learning (DRL) with CF, employing the actor–critic method and the Deep Deterministic Policy Gradient (DDPG) algorithm. This integration enhances the system’s ability to navigate the recommendation space effectively, especially for new users with less interaction data. The system utilizes DRL for making initial recommendations to new users and to generate optimal recommendation as more data becomes available. Additionally, singular value decomposition (SVD) is used for matrix factorization in CF, improving the extraction of detailed embeddings that capture the latent features of users and movies. This approach significantly increases recommendation precision and personalization. Our model’s performance is evaluated using the MovieLens dataset with metrics like Precision, Recall, and F1 Score and demonstrates its effectiveness compared with existing recommendation benchmarks, particularly in addressing sparsity and new-user challenges. Several benchmarks of existing recommendation models are selected for the purpose of model comparison. Full article

Despite the advances in computational literary analysis of Western literature, in-depth analysis of the South Asian literature has been lacking. Thus, social network analysis of the main characters in the Indian epic Mahabharata was performed, in which it was prepossessed into verses, followed by a term frequency–inverse document frequency (TF-IDF) transformation. Then, Latent Semantic Analysis (LSA) word vectors were obtained by applying compact Singular Value Decomposition (SVD) on the term–document matrix. As a novel innovation to this study, these word vectors were adaptively converted into a fully connected similarity matrix and transformed, using a novel locally weighted K-Nearest Neighbors (KNN) algorithm, into a social network. The viability of the social networks was assessed by their ability to (i) recover individual character-to-character relationships; (ii) embed the overall network structure (verified with centrality measures and correlations); and (iii) detect communities of the Pandavas (protagonist) and Kauravas (antagonist) using spectral clustering. Thus, the proposed scheme successfully (i) predicted the character-to-character connections of the most important and second most important characters at an F-score of 0.812 and 0.785, respectively, (ii) recovered the overall structure of the ground-truth networks by matching the original centralities (corr. > 0.5, p < 0.05), and (iii) differentiated the Pandavas from the Kauravas with an F-score of 0.749. Full article