Search Results (21,376)

The increasing complexity of next-generation wireless systems challenges the scalability and generalization capabilities of traditional model-driven physical layer (PHY) design, which relies on analytically derived channel models and optimization frameworks. This paper presents a comprehensive survey and critical review of deep learning (DL) architectures enabling the transition toward AI-native PHY design. A unified optimization perspective is developed in which all PHY tasks—including channel estimation, channel state information (CSI) feedback, massive MIMO processing, signal detection, channel coding, beamforming, resource allocation, and semantic-aware transmission—are formulated under a common empirical risk minimization (ERM) framework. Neural architectures such as autoencoders, convolutional and recurrent networks, transformers, and reinforcement learning models are examined through their underlying optimization formulations, loss functions, training methodologies, and representation learning mechanisms. The review compares model-driven and AI-native approaches in terms of performance metrics, computational complexity, robustness, generalization capability, and practical deployment constraints, including hardware limitations, energy efficiency, and real-time feasibility. The analysis highlights the conditions under which AI-native architectures provide adaptability and performance improvements while identifying trade-offs in complexity, latency, and interpretability. The study concludes by outlining prioritized research directions toward fully adaptive and self-optimizing wireless communication systems. Full article

Transparent clinical decision-making remains a critical barrier to deploying deep learning in medical diagnosis. Post hoc explanation methods approximate model behaviour after training but cannot guarantee that explanations faithfully reflect the underlying reasoning. This study proposes a Self-Explaining Neural Network (SENN) for Parkinson’s Disease (PD) screening via Ground Reaction Force (GRF) gait analysis, enforcing intrinsic interpretability through learnable basis concepts and input-dependent relevance scores computed jointly with the prediction. The architecture combines a four-block residual CNN backbone with stochastic depth regularisation, a 16-concept encoder with diversity and stability constraints, and temperature-scaled probability calibration for reliable clinical operating points. Evaluated on the PhysioNet Gait in Parkinson’s Disease dataset (306 subjects, 16 GRF sensors per foot), SENN achieves a subject-level ROC-AUC of 0.916 [95% CI: 0.867–0.964], sensitivity of 0.913 [0.862–0.963], specificity of 0.671 [0.485–0.858], and Average Precision of 0.942 [0.918–0.967], reported across five independent random seeds. Comparative evaluation against four deep learning baselines—CNN-Residual, BiLSTM, CNN-LSTM, and CNN-Attention—confirms that the interpretability constraints impose no statistically significant reduction in discriminative performance, with all pairwise ROC-AUC confidence intervals overlapping. Concept-level analysis reveals that the three most discriminative concepts correspond to disrupted midfoot loading patterns, increased step-length variability, and bilateral cadence asymmetry—all established biomechanical hallmarks of parkinsonian gait—providing clinically grounded, patient-specific explanations without post hoc approximation. These findings demonstrate that rigorous intrinsic interpretability and competitive predictive accuracy are simultaneously achievable in deep gait analysis, supporting the clinical adoption of transparent diagnostic AI. Full article

Most neural architectures model time as a one-dimensional real-valued variable, constraining temporal reasoning to sequential propagation along a single axis. We introduce Complex-Time Neural Networks (CTNN), a new class of architectures in which temporal coordinates are elements of the complex plane T = t + iτ ∈ ℂ, where Re(T) preserves chronological ordering and Im(T) encodes an orthogonal experiential dimension. Within this geometry, Im(T) < 0 defines a memory domain enabling retrospective retrieval, Im(T) = 0 corresponds to present-moment computation, and Im(T) > 0 defines an imagination domain for prospective projection. We prove the Expressive Separation Theorem (Theorem 1), establishing that, within the temporally coupled function class G_TCP and under explicit Assumptions A1–A4 (in particular the bounded projection Assumption A3), CTNN accesses temporally coupled functions at O(1) cost with respect to temporal distance Δ₁, Δ₂, while real-time architectures incur Ω(Δ₁ + Δ₂) sequential steps. For layered compositions, this yields an exponential composition gap within G_TCP under A1–A4. These advantages hold under the stated assumptions and may not directly generalize to broader function classes or large-scale settings where A3 cannot be maintained. Therefore, Theorem 1 provides a formal separation result for G_TCP, while CTNN more broadly defines a geometric framework for temporal computation. As the first concrete instantiation of this framework, we develop Complex-Time Convolutional Neural Networks (CTCNN). CTCNN achieves state-of-the-art performance on Something-Something V2 (70.2 ± 0.4%, +1.1% over VideoMAE v2, p < 0.01), strong performance on Kinetics-400 (78.4 ± 0.3%), and substantial gains on Long Range Arena Path-X (87.3% vs. 79.6%, +7.7%), using 3.4× fewer parameters than VideoMAE v2. Learnable angular parameters α and β provide computationally interpretable parameters related to memory-access span and prospection breadth, with values varying systematically across task families. Full article

Owing to the intricate mechanical coupling characteristics and the considerable difficulty in extracting synergistic spatio-temporal features from high-dimensional sensor data under fluctuating alternating loads, this study proposes a robust anomaly detection framework that combines Normalized Mutual Information (NMI) and Spatio-Temporal Graph Neural Networks (STGNN). First, NMI is utilized to quantify the nonlinear physical coupling intensity among multi-source sensors, thereby filtering out weakly correlated noise and reconstructing the spatial topological structure of the coupler system. Subsequently, a deep learning architecture incorporating Graph Convolutional Networks (GCN), Gated Recurrent Units (GRU), and one-dimensional convolutional residual connections is developed to capture the dynamic evolutionary characteristics of equipment states across both spatial interactions and temporal sequences. Finally, based on the model’s health-state predictions, a moving average algorithm is introduced to smooth the residual sequences, and an anomaly early-warning baseline is established in conjunction with the 3σ criterion. Experimental validation conducted using field service data from heavy-haul trains demonstrates that, compared to conventional serial CNN and Long Short-Term Memory (LSTM) models, the proposed method exhibits superior fitting performance and robustness against noise, effectively reducing the false alarm rate within normal working intervals. In a real-world case study, the method successfully identified variations in spatial linkage features induced by local damage and triggered timely alerts. Notably, the spatial alarm nodes were highly consistent with the fatigue crack initiation sites identified through on-site magnetic particle inspection. This study provides a viable data-driven analytical framework for the condition monitoring and anomaly identification of critical load-bearing components in heavy-haul trains. Full article

The Industrial Internet of Things (IIoT) has become a cornerstone of modern industrial automation, enabling real-time monitoring, intelligent decision-making, and large-scale connectivity across cyber–physical systems. However, the growing scale, heterogeneity, and dynamic behavior of IIoT environments significantly expand the attack surface and challenge the effectiveness of conventional security mechanisms. In this paper, we propose DeepSense, a hybrid and adaptive anomaly and intrusion detection framework specifically designed for resource-constrained and heterogeneous IIoT deployments. DeepSense integrates three complementary components: DataSense, a realistic data pipeline and experimental testbed supporting synchronized sensor and network data processing; RuleSense, a lightweight rule-based detection layer that provides fast, deterministic, and interpretable anomaly screening at the edge; and NeuroSense, a learning-driven detection module comprising an adaptive ensemble of 22 machine learning and deep learning models spanning classical, neural, hybrid, and Transformer-based architectures. NeuroSense operates as a second detection stage that validates suspicious events flagged by RuleSense and enables both coarse-grained and fine-grained attack classification. To support rigorous and practical assessment, this work further introduces a comprehensive performance evaluation framework that extends beyond accuracy-centric metrics by jointly considering detection quality, latency, resource efficiency, and detection coverage, alongside an optimization-based process for selecting Pareto-optimal model ensembles under realistic IIoT constraints. Extensive experiments across diverse detection scenarios demonstrate that DeepSense exhibits strong generalization, lower false positive rates, and robust performance under evolving attack behaviors. The proposed framework provides a scalable and efficient IIoT security solution that meets the operational requirements of Industry 4.0 and the resilience-oriented objectives of Industry 5.0. Full article

This work explores the application of Spiking Neural Networks (SNNs) and Continual Learning (CL) methodologies to the problem of steering angle regression, using autonomous driving simulation as the experimental context, with a focus on energy efficiency and alignment with sustainable computing objectives. The primary goal was to design and implement CL techniques in SNNs to assess the model’s ability to maintain accuracy in explored environments while reducing CO${}_2$ emissions through the optimized use of a subset of the data. This study emerges in response to the increasing energy demand of deep learning models, which poses a challenge to sustainability. SNNs, inspired by the efficiency of biological neural systems, offer significant advantages in terms of computational and energy consumption, making them a promising alternative. CL techniques, such as Elastic Weight Consolidation and replay memory, are integrated to mitigate catastrophic forgetting in sequential learning tasks. The methodology includes adapting the PilotNet architecture for SNNs, preprocessing datasets generated in the Udacity driving simulator, and evaluating models in incremental learning scenarios. The experiments compare the performance of SNNs with CL against baseline models without CL, using mean squared error (MSE), computational efficiency, and equivalent CO${}_2$ emissions as evaluation metrics. The results demonstrate that replay memory enables the retention of prior knowledge with a limited increase in energy consumption. This work concludes that SNNs with CL are a viable alternative for sustainable AI applications. Future research directions include a focus primarily on hardware-specific implementations and real-world testing. Full article

Molten salt chlorination is a key industrial route for producing titanium tetrachloride (TiCl₄), yet the atomistic catalytic role of iron (Fe) in the carbothermic chlorination of titanium oxides remains unclear. Here, the chlorination behavior of the NaCl–C–Cl₂–FeTiO₃ system was investigated by combining thermodynamic calculations with Ab Initio Molecular Dynamics (AIMD) and Deep Potential Molecular Dynamics (DPMD) simulations. AIMD results show that carbon adjacent to Fe exhibits enhanced reactivity, and that Fe-C synergistic electron transfer promotes both titanium oxide reduction and subsequent titanium chlorination. DPMD results further reveal that Fe not only accelerates these transformations, but also improves interfacial contact among carbon, titanium oxides, and molten salt, thereby enhancing mass transfer and shortening the formation time of TiCl₄. Temperature-dependent analysis indicates that Fe-C and C-O coordination numbers remain high near 1073 K, where TiCl₄ formation is efficient and relatively stable. Although increasing temperature can further enhance diffusion, its effect on reaction acceleration is limited, while excessively high temperatures weaken Fe-C interactions and reduce catalytic efficiency. These findings clarify the catalytic mechanism of Fe in molten salt chlorination at the atomic scale and provide theoretical support for process optimization. Full article

Winter pruning is a key management practice in viticulture that directly affects vine architecture, yield balance, and grape quality. At the same time, it is a highly labor-intensive operation, and the selective identification of appropriate cutting locations remains one of the main challenges limiting the automation of pruning in vineyards. Advances in machine vision provide new opportunities to support the development of robotic pruning systems. The objective of this study was to develop and evaluate a vision-based method for estimating grapevine pruning points and cutting lines using instance segmentation outputs generated by YOLO models. A dataset of 1500 RGB images of dormant grapevines was collected under field conditions in the Nobilis vineyard located in southeastern Poland. Two annotation strategies were implemented to define pruning regions. YOLO-based instance segmentation models were trained and evaluated for detecting cutting-related structures. Based on the predicted segmentation masks, a geometry-based method termed PCAcutSeg-V was developed to estimate class-dependent cutting points and cutting lines using principal component analysis applied to object contours. The results indicate that YOLOv8 and YOLO11 architectures achieved the highest segmentation performance among the evaluated models. The simplified annotation strategy provided more stable geometric inputs for the PCAcutSeg-V method, enabling more reliable estimation of cutting points and cutting lines compared with the extended annotation approach. When combined with the PCAcutSeg-V method, the proposed perception–geometry pipeline achieved high effectiveness in pruning decision estimation. The method was further implemented in a real-time processing pipeline using an RGB camera and an edge computing platform, where it maintained performance consistent with the results obtained from offline image analysis. These findings demonstrate that combining deep learning-based instance segmentation with deterministic geometric reasoning enables accurate and interpretable estimation of grapevine pruning locations and provides a promising foundation for future autonomous pruning systems. Full article

The growing complexity of unwanted messages, especially SMS spam, presents a serious challenge to the security of digital communication and user experience. While conventional spam detection models are useful on clean datasets, they are vulnerable to targeted attacks that aim to evade detection. This study is motivated by the urgent need to evaluate the resilience of machine learning models against evolving threats in real-world applications. We specifically investigate the robustness and interpretability of a Multinomial Naive Bayes (MNB) model, representative of traditional machine learning, and a Tiny Text convolutional neural network (Tiny Text CNN), representative of deep learning models, for SMS spam detection. Using the UCI dataset under simulated adversarial text attacks, both models were tested against filler-word insertion and character-level perturbation attacks. Results show that while the Tiny Text CNN maintained higher overall robustness (accuracy: 0.9821 clean vs. 0.9758 under character attacks), both models experienced notable degradation in recall, with MNB being more susceptible to filler-word attacks. Interpretability analyses using LIME and gradient-based saliency maps indicated that adversarial perturbations alter feature importance, diminishing the influence of spam-indicative tokens. The findings underscore the trade-offs between model complexity and adversarial resilience, offering insights for developing more secure and interpretable spam detection systems. Full article

Deep neural networks (DNNs) are traditionally analyzed as black-box function approximators, yet their internal structure exhibits phase transitions characteristic of complex physical systems. In this study, we investigate topological–functional decoupling—the phenomenon whereby a network retains full graph connectivity while losing computational function—in trained neural networks through the lens of percolation theory. By subjecting three distinct architectures (Shallow, Deep, and Wide MLPs) to a unified edge-pruning analysis on Fashion-MNIST, we uncover a fundamental divergence between structural integrity and computational capacity in this experimental setting. We report three key phenomena observed in these experiments: (1) the zombie network state under stochastic pruning, where the system retains global connectivity ($P_{\infty }\approx 1.0$) yet suffers a catastrophic functional collapse (accuracy falls below 50% of baseline at prunning ratio $p_f\approx 0.35$–$0.68$ depending on depth), proves that graph reachability does not imply computational capability; (2) depth fragility, where increased network depth triggers multiplicative signal decay (the avalanche effect), rendering deep architectures exponentially more vulnerable to random edge removal than shallow ones ($p_f^{\mathrm{deep}}\approx 0.35$ vs. $p_f^{\mathrm{shallow}}\approx 0.68$); and (3) scale-free universality, observed under magnitude-based pruning, where a robust functional skeleton maintains accuracy near the baseline (∼89%) up to extreme sparsity ($p_f\approx 0.85$–$0.95$) across all three architectures. Robustness stems not from holographic redundancy in the overall connection count but from the emergent heavy-tailed rich-club organization of weight magnitudes—a sparse set of high-magnitude synapses that form the functional backbone of the network, decoupled from the redundant topological mass. These findings offer new physical constraints for the design of resilient neuromorphic hardware. Full article

Deep learning (DL) has transformed cancer imaging by enabling automated tumour detection, classification, and risk prediction. Despite impressive diagnostic performance, limited explainability and poor uncertainty calibration continue to restrict clinical integration. This review is guided by five research questions that examine the challenges, impact, and translational implications of explainable artificial intelligence (XAI) in oncology imaging. We identify key barriers to trust, including dataset bias, shortcut learning, opacity of convolutional neural networks, and workflow misalignment. Evidence suggests that explainable models can increase clinician confidence, reduce false positives, and improve collaborative decision-making when explanations are faithful, semantically meaningful, and uncertainty aware. We evaluate architectural strategies that embed interpretability such as concept-bottleneck models, prototype-based learning, and attention regularization along with post hoc techniques. Beyond performance metrics, we examine how interpretable AI aligns with clinical reasoning processes and analyse regulatory, ethical, and medico-legal considerations influencing deployment. The findings indicate that explainability alone is insufficient, durable trust requires epistemic alignment, prospective validation, lifecycle governance, and equity-focused evaluation. By reframing explainability as a structural design principle rather than a supplementary feature, this review outlines a pathway toward accountable and clinically dependable AI systems in oncology. Full article

Remote sensing change detection (RSCD), a fundamental task in Earth observation, aims to automatically identify land-cover changes (e.g., building construction, vegetation degradation) by comparing multitemporal satellite or aerial images of the same region. With the explosive growth of high-resolution remote sensing data, achieving real-time accurate change detection on edge computing devices (e.g., drone-embedded chips, satellite on-board processors) has become an urgent challenge—existing deep learning methods, despite high accuracy, are hindered by massive parameters and computational costs that preclude deployment on resource-constrained embedded hardware. To address this, we focus on lightweight (i.e., low parameter count and low computational cost) RSCD network design, targeting three critical bottlenecks: blurred boundaries of changed regions, missed detection of small objects, and insufficient computational efficiency. We propose ATCFNet (Adjacent-Temporal Cross Fusion Network), featuring a three-step progressive feature optimization strategy: (1) the Adjacent Feature Aggregation Module (AFAM) enhances shallow geometric details via lateral three-stage fusion to compensate for lightweight backbones; (2) the Temporal Attention Cross Module (TACM) integrates cross-level feature propagation and Convolutional Block Attention Module (CBAM) for collaborative optimization of high-level semantics and low-level details; and (3) the Efficient Guidance Module (EGM) establishes long-range dependencies using shared change priors and lightweight self-attention to suppress internal voids in changed regions. Experiments on three public datasets (LEVIR-CD, HRCUS, SYSU-ChangeDet) demonstrate that ATCFNet achieves state-of-the-art accuracy with merely 3.71 million (M) parameters and 3.0 billion (G) floating-point operations (FLOPs)—F1-scores of 91.46%, 77.05%, and 83.53%, significantly outperforming 18 existing methods in most indicators. Notably, it excels in edge integrity (avoiding jagged blurring at change boundaries) and small-target detection in high-resolution urban scenes. This study provides an efficient and reliable lightweight solution for edge computing scenarios such as real-time drone inspection and satellite on-board intelligent processing. Full article

Left-turn lanes are critical elements of urban intersections. Accurate and efficient lane detection is essential for the safe navigation of autonomous vehicles. To address the limitations of existing semantic segmentation algorithms—specifically, inadequate detection accuracy, high computational cost, and vulnerability to environmental disturbances—we propose a lightweight deep convolutional neural network named EA-UNet. First, we replace the standard U-Net encoder with EfficientNet-B0 to enhance feature extraction efficiency. Second, we introduce a novel contextual coordination module, termed MP-ASPP, which integrates a Convolutional Block Attention Module (CBAM) to further refine attention mechanisms. Finally, a comprehensive real-world dataset was constructed by collecting videos and images of left-turn waiting areas during real-vehicle testing. Experimental results demonstrate that EA-UNet significantly outperforms the baseline U-Net and other state-of-the-art models, achieving accurate and efficient segmentation of left-turn lanes even in complex scenes. Full article

Investigating soil seepage considering free surface conditions under complex geological conditions is of great significance to ensure the safety of dams. In recent years, physics-informed deep learning (PINN) has become a cross-disciplinary hotspot for solving forward and inverse problems based on partial differential equations. However, the challenges in free surface simulation have confined the majority of current PINN research to seepage problems under fixed boundary conditions. To address the above issues, we propose a physics-informed deep learning-based approach for steady seepage in dams. In the proposed method, two different free surface approximation strategies are introduced to accommodate varying boundary conditions in the dam seepage problem. The first strategy approximates the free boundary by sampling points, while the second strategy approximates the free boundaries by an additional deep neural network. To validate the proposed methods, three benchmark cases with different boundary conditions have been conducted. The results indicate that the proposed approach effectively simulates steady seepage in dams. Both point-sampling and deep neural network-based free surface approximation strategies demonstrate high accuracy in predicting the location of the phreatic surface and the discharge of the seepage. Specifically, the prediction results are comparable in accuracy to analytical solutions and advanced numerical simulation methods. Full article

Many place recognition approaches, which identify previously visited places or locations by matching current sensory data, such as 2D RGB images and 3D point clouds, have been proposed to achieve accurate and robust localization and loop closure detection in global positioning system (GPS)-denied environments. Since visual place recognition (VPR) methods that rely on images captured by camera sensors are highly sensitive to variations in appearance, including changes in lighting, surface color, and shadows, they can lead to poor place recognition accuracy. In contrast, light detection and ranging (LiDAR)-based place recognition (LPR) approaches based on 3D point cloud data that captures the shape and geometric structure of the environment are robust to changes in place appearance and can therefore provide more reliable place recognition results than VPR methods. This work presents an indoor LPR method called PointNetVLAD-based indoor pedestrian localization (PIPL). PIPL is a deep network model that uses PointNetVLAD to learn to extract global descriptors from 3D LiDAR point cloud data. PIPL can recognize places previously visited by a pedestrian using point clouds captured by a low-cost LiDAR sensor on a smartphone in small-scale indoor environments, while PointNetVLAD performs place recognition for vehicles using high-cost LiDAR, GPS, and inertial measurement unit (IMU) sensors in large-scale outdoor areas. For place recognition on 3D point cloud reference maps generated from LiDAR scans, PointNetVLAD exploits the universal transverse mercator (UTM) coordinate system based on GPS and IMU measurements, whereas PIPL uses a virtual coordinate system designed in this study due to the unavailability of GPS indoors. In experiments conducted in campus buildings, PIPL shows significant advantages over NetVLAD (known as a convolutional neural network (CNN)-based VPR method). Particularly in indoor environments with repetitive scenes where geometric structures are preserved and image-based appearance features are sparse or unclear, PIPL achieved $39\%$ higher top-1 accuracy and $10\%$ higher top-3 accuracy compared to NetVLAD. Furthermore, PIPL achieved place recognition accuracy comparable to NetVLAD even with a small number of points in a 3D point cloud and outperformed NetVLAD even with a smaller model training dataset. The experimental results also indicate that PIPL requires over $76\%$ less place retrieval time than NetVLAD while maintaining robust place classification performance. Full article