Search Results (7,971)

We propose an end-to-end differentiable framework for hedonic agricultural land pricing that integrates deep learning-based land cover classification with spatial econometric modeling into a single neural architecture. Traditional hedonic pricing approaches typically separate land cover extraction from price regression, leading to suboptimal feature representations that fail to capture the spatial spillover effects inherent to agricultural markets. In our system, a Swin Transformer-based semantic segmentation network extracts pixel-level land cover features from high-resolution multispectral imagery, which are then aggregated within parcel boundaries to produce composition vectors. These features are combined with static parcel attributes and fed into a graph isomorphism network that models spatial dependencies among neighboring parcels through message passing. The central methodological innovation is a differentiable Moran’s I operator that computes spatial autocorrelation from predicted parcel prices and incorporates this statistic into the training objective as a regularizing loss term. This constraint explicitly penalizes deviations from empirically observed target levels of positive spatial autocorrelation in agricultural land markets, thereby ensuring that the learned land cover features are optimized to explain spatial price clustering rather than generic class categories. The complete pipeline, including the segmentation backbone, graph neural network, and spatial autocorrelation computation, is fully differentiable, allowing gradients from the spatial loss to flow backwards and update pixel-level features. This design transforms land cover classification from a mere preprocessing step into an economically informed feature-learning process. The unified framework thereby produces parcel valuations that are both pixel-accurate and spatially coherent, capturing complex nonlinear dependencies such as irrigation network effects or soil-type continuity that conventional spatial econometric models cannot represent. By jointly optimizing segmentation features and their spatial spillover effects on market prices, our approach represents a significant departure from the two-stage hedonic pricing methodology. Full article

Although recommendation systems based on neural collaborative filtering (NCF) can achieve high rating accuracy, the point estimates they generate do not express prediction confidence. Existing variants that account for uncertainty, such as Monte Carlo dropout, deep ensemble models, and single-head Bayesian networks, can only capture epistemic uncertainty. We propose BayesNCF-v2, a single-model Bayesian neural collaborative filtering framework that combines a Bayes-by-Backprop output layer with a heteroscedastic head trained under the Gaussian NLL, thereby jointly learning aleatoric and epistemic uncertainty. Across five seeds and three datasets, compared to standard NCF, it reduces ECE by 25–73% and NLL by 7–20%, while its calibration performance shows no statistically significant difference from heteroscedastic MC dropout and approaches that of a heteroscedastic deep ensemble consisting of five models, despite having significantly fewer parameters. For cold-start users, the learned aleatoric variance increases, thereby reducing the ECE for cold-start users on three datasets. These results hold consistently across both random and temporal partitions, demonstrating that heteroscedastic aleatoric modeling is the main driver of calibration in these experiments. Full article

In Santiago, Chile, the 8 h average atmospheric ozone (O₃) concentration frequently exceeds the national air quality standard of 61 ppbv (parts per billion by volume). Due to local meteorological and topographic conditions, the highest ozone levels are observed at the Las Condes station in eastern Santiago. To provide a reliable tool for forecasting adverse air quality events, we developed two predictive models for the daily maximum 8 h average concentration: an XGBoost decision tree model and an attention-based Long Short-Term Memory (LSTM) deep neural network. Given the photochemical coupling between ozone and nitrogen oxides, models were also implemented to forecast daily concentrations of nitric oxide (NO) and nitrogen dioxide (NO₂). Models were trained on data spanning 2007–2022 and evaluated on a strictly out-of-sample period (January 2023 to July 2024). The attention-based LSTM achieved symmetric mean absolute percentage errors (sMAPE) of 8.2%, 21%, and 17% for O₃, NO, and NO₂, respectively. It outperformed XGBoost across all eight statistical metrics evaluated, while both models demonstrated substantial improvements over a previous-day persistence baseline. Full article

Catastrophic forgetting is a major challenge for deep learning models when they are incrementally trained on a sequence of new data. Reducing this forgetting in image and video data has been the primary research focus, but less attention has been given to textual domains, where discrete token distributions and semantic shifts occur across different topics. Furthermore, standalone strategies proposed for reducing catastrophic forgetting still have room for improvement. To this end, this paper proposes a synergy of stratified memory replay with parameter regularization for a BiLSTM-based incremental learning model to mitigate catastrophic forgetting in sequential text datasets. The stratified replay mechanism replays a small buffer of historical data samples into current training phases to preserve the old data patterns, while the parameter regularization penalizes modifications to the neural network weights crucial to the past tasks. The proposed approach is evaluated in an incremental training pipeline on distinct textual datasets, including software bug reports (Task A), a news dataset (Task B), and emails (Task C). The evaluation results demonstrate that the baseline neural network experiences catastrophic forgetting as its initial dataset (Task A) accuracy drops from $87.53\%$ to $11.75\%$. The standalone experience replay approach manages to retain Task A accuracy at $80.01\%$, down from its peak of $86.56\%$, while for Task B, it achieves $93.97\%$, down from the peak of $98.37\%$. The buffer sensitivity analysis indicates the model accuracy improves with increasing replay buffer size. The parameter regularization approach effectively reduces catastrophic forgetting, but it remains less effective for disruptive text distribution sequences, resulting in noticeable forgetting on prior tasks and reduced plasticity on later tasks. Evaluations of this approach show that Task A accuracy is reduced from $87.87\%$ to $71.09\%$ after training on Task C. The proposed hybrid approach reduces forgetting and preserves Task A and Task B accuracies at $83.73\%$ and $95.04\%$, respectively. Thus, the empirical evaluations demonstrate that the proposed approach is effective and outperforms the standalone experience replay strategy and parameter regularization, limiting the forgetting on the earliest task to just $4.65\%$ compared to $6.55\%$ forgetting of the replay-based method, while allowing enough plasticity for the final task to reach $98.92\%$ accuracy. Full article

The sponge construction serves as a fundamental design framework for hash functions and authenticated encryption algorithms. In differential cryptanalysis of large-state permutations underlying such algorithms, conventional approaches are constrained by state size and diffusion speed. The feature extraction capability of deep learning offers a potential alternative to mitigate these limitations. To improve the distinguishing performance of differential neural distinguishers against sponge-based algorithms, a methodology integrating data construction and network architecture optimization is proposed. Specifically, a multi-sample triplet input format is designed to preserve differential characteristics, and a convolutional block attention module is introduced to capture long-range dependencies along both the channel and spatial dimensions within the large-state permutation. Experimental evaluations were conducted on the Keccak and Ascon algorithms. For Keccak, the maximum distinguishable round number was identified as 3. At this round number, Keccak-p achieved full distinguishability (100% accuracy), while the sponge-based SHA-3-256 attained a distinguishing accuracy of 99.99%, improving upon the previous best result by 0.95 percentage points. For Ascon, the maximum distinguishable round number was 4, where Ascon-p achieved an accuracy of 54.85% with 64 sample pairs—the highest reported accuracy for this setting—while delivering comparable performance at the matched 32-pair setting (53.40% vs. 53.54% in prior work) with approximately one-twelfth of the training epochs; under the same setting, the sponge-based Ascon-Hash256 achieved an accuracy of 53.06%. These findings demonstrate the effectiveness of the proposed framework in enhancing neural distinguisher accuracy against sponge-based algorithms and offer an analytical approach for empirical security evaluation, with results qualitatively consistent with the indifferentiability bound of the sponge construction. Full article

Introduction: Voice-based digital biomarkers have emerged as a promising, non-invasive approach for the early detection and monitoring of neurodegenerative disorders, particularly Parkinson’s disease (PD). Although voice recordings can be acquired easily using mobile health technologies, their integration into routine clinical practice remains limited due to challenges related to model interpretability and clinical validation. This study aimed to investigate the diagnostic potential of voice recordings acquired through the TALIA smartphone and web-based platform and to improve model transparency using explainable artificial intelligence (XAI) techniques. Methods: Voice recordings from participants with PD and control group (CG) were acquired over a six-month period using the TALIA digital-health platform at Vito Fazzi Hospital in Lecce, Italy. The data were evaluated cross-sectionally at the recording level. Sustained vowel (/a/) phonations were preprocessed and transformed into spectrogram images. A transfer-learning framework based on the pre-trained VGGish convolutional neural network (VGGish-CNN) was developed to classify PD and CG voice recordings. To enhance interpretability, explainable artificial intelligence (XAI) methods, including Local Interpretable Model-Agnostic Explanations (LIME) and Occlusion Sensitivity, were integrated to identify the spectro-temporal regions contributing most strongly to model predictions. Results: The proposed VGGish-CNN framework demonstrated excellent classification performance in distinguishing PD from CG. The model achieved an accuracy of 0.93, precision of 0.91, recall of 0.95, F1-score of 0.93, loss of 0.16, and an area under the receiver operating characteristic curve (AUC) of 0.987. XAI analyses provided qualitative, sample-specific visualizations of the spectro-temporal regions contributing to individual model predictions, thereby improving the transparency and interpretability of the deep-learning model. Conclusions: The findings demonstrate that a transfer-learning approach based on VGGish-CNN could achieve promising recording-level classification performance in distinguishing voice recordings from participants with PD and CG within this pilot dataset. Furthermore, XAI techniques provide qualitative insight into the model’s decision-making process. These preliminary findings support further investigation of voice-based approaches for PD screening in larger cohorts using participant-level and external validation. Full article

Deep neural networks have achieved strong performance in medical image classification, but their deployment may be constrained by the computational cost of high-capacity models. Knowledge distillation (KD) addresses this problem by transferring knowledge from a teacher to a lightweight student. However, the reliability of teacher supervision may vary across samples, potentially introducing noisy guidance and local conflicts with ground-truth supervision. We propose Selective Confidence-guided Projection-based Encoding (SCOPE), a conflict-aware KD framework comprising Selective Relation Alignment (SRA) and Gradient Conflict Resolution (GCR). SRA constructs reliability-aware relational supervision by combining teacher-derived relations with dataset-specific auxiliary priors, whereas GCR removes distillation-gradient components that conflict with the classification objective. Experiments on nine medical image datasets and multiple teacher–student architectures demonstrate competitive predictive performance, improved training stability, and low computational overhead. Full article

Geospatial artificial intelligence (GeoAI) integrates multimodal remote sensing with deep learning to model complex agricultural systems at scale. Within this framework, accurate and non-destructive prediction of seed composition from in-season standing crops is essential for breeding and precision agriculture. This study developed an end-to-end convolutional neural network (CNN) framework to estimate eight seed traits (protein, oil, sucrose, fiber, starch, ash, complex and simple carbohydrates) from UAV-based multisensor imagery and associated genotype and phenological metadata. A total of 372 soybean samples were collected over two growing seasons (2020–2021) from two fields in Missouri, with UAV flights capturing multispectral (MSI), thermal (THR), and LiDAR (LDR) data at four time points spanning vegetative to reproductive growth stages. CNN models were trained in single- and multi-date configurations, incorporating genotype (GEN) and days after sowing (DAS) as additional features. The highest accuracy was achieved for sucrose (R² = 0.80), followed by simple carbohydrate (R² = 0.70) and starch (R² = 0.55), with notable gains from GEN and DAS. Multi-date models incorporating earlier acquisitions often matched or outperformed later or all-date combinations. Among modalities, MSI provided the most robust estimates, with limited added value from LDR or THR. Unlike feature-based pipelines prone to multicollinearity, this image-to-trait approach enables automated, scalable prediction of soybean seed composition for in-season, field-level assessment. Full article

A brain tumor is a life-threatening disease that carries a high mortality burden. It can be treated if detected in the early stages. Manual analysis by a radiologist using Magnetic Resonance Imaging (MRI) is effective but slow and subject to inter-observer variability. In order to automate it, various deep learning models, especially Convolutional Neural Networks (CNNs), have been widely used. However, most of them rely on frozen, ImageNet-based pretrained backbones, where only a small classification head is updated during training. This leaves a large segment of parameters unable to adapt to the domain shift between natural images and MRI scans. This limits how far such models can be modified or improved further. This makes the model highly parameter-centric and unsuitable for development. In addition, most studies fail to provide insights on architectural choices, which inhibits reproduction and modifications. In order to alleviate these issues, this study presents SERA-Net, a convolutional architecture for brain tumor classification using MRI scans. It is trained from scratch and quantifies the performance of each model using an incremental ablation study. It consists of four blocks comprising different components, such as Squeeze-and-Excite channel attention, residual connections, a wider classification head, etc. Rather than proposing new operations, the contribution lies in the systematic, ablation-driven combination and empirical validation of established architectural components for this specific task. The model was trained and evaluated on a publicly available MRI dataset. With just ≈5 million parameters, it achieved an accuracy of 95.44%, an F1-score of 0.953, and various near-ideal parameters. It was observed to outperform six pretrained CNN baselines by more than 5–10%, while using substantially fewer parameters. Various other analyses indicate the efficacy of the proposed approach in detecting brain tumors. The proposed approach provides a scalable design that can be adapted to different domains with limited modifications. Full article

Accurately measuring geometric parameters of high-aspect-ratio blind holes is a critical requirement for microelectronics and micro–nanooptics applications. Conventional measurement techniques are restricted by limited measurement precision, incomplete measurable geometric indicators, bulky optical hardware, and potential irreversible damage to test specimens. To tackle these drawbacks, this paper proposes a deep-learning-driven image reconstruction method for full-parameter geometric measurement of high-aspect-ratio blind holes. In the proposed method, a Denoising Convolutional Neural Network (DnCNN) is deployed to restore low signal-to-noise ratio (SNR) microscopic images captured at blind hole bottoms. Meanwhile, Laplacian variance sharpness evaluation coupled with local quadratic polynomial fitting is adopted to boost the precision of Z-axis focal positioning and depth calculation. A dedicated machine vision measurement system is custom-developed, where annular ring illumination is integrated to strengthen light irradiation at hole bottoms and guarantee high-quality image acquisition. To address the scarcity of authentic paired clean-noisy training data, high-clarity surface micrographs are artificially degraded in reverse to synthesize low-quality counterparts, forming a dataset containing 2000 image pairs for DnCNN training and optimization. Furthermore, a dedicated feature extraction pipeline combining image preprocessing, Otsu-based adaptive threshold segmentation and least-squares ellipse fitting is designed to extract core geometric metrics, including microhole diameter, depth, taper angle, and relative ellipticity deviation. Experimental validation reveals that the absolute measurement deviations of the mean measured top diameter, bottom diameter, and depth reach 4 μm, 7 μm, and 8 μm, respectively, with a calculated depth-to-diameter-ratio error of 0.06. The results demonstrate that this method delivers a high-precision, low-cost non-destructive micrometer-scale solution for comprehensive geometric inspection of high-aspect-ratio blind holes. Full article

Deep neural networks (DNNs) are widely deployed in safety-critical applications such as medical diagnosis and autonomous driving. Adversarial training (AT) is among the most effective defenses, casting robust optimization as a min–max problem over a defender-specified ${\ell }_p$-ball of fixed radius $ϵ$. Bounded defenses of this kind are known to generalize poorly to test-time perturbations larger than $ϵ$. In this paper, we revisit this failure mode on MNIST and FashionMNIST and make three of its properties explicit. The degradation is abrupt rather than gradual, and its location is indexed by the training budget, so that enlarging the budget translates the drop instead of removing it. The translation is, in turn, capped by trainability, since training stops converging once $ϵ$ becomes too large. Decision-surface visualization exhibits the same failure geometrically, as adversarially trained models form a plateau whose edge coincides with the training boundary. Together, these properties suggest that the failure follows from concentrating training on a single radius. We therefore present Density-Weighted Adversarial Training (DWAT), a plug-in framework that spreads training over a set of sampled $ϵ$-balls and reweights each candidate adversarial example by a Gaussian density of its distance from the benign sample. We derive its objective as a self-normalized importance-sampling estimate of an expected adversarial risk under a perturbation prior, and we show that this risk is upper-bounded by the standard adversarial risk, so that the in-bound objective of DWAT relaxes rather than replaces that of AT. We further prove that the population objective upper-bounds the worst-case adversarial risk at every radius, including radii beyond the training budget, at an explicit cost that grows with the radius. Experiments with PGD-AT, TRADES, and MART as base defenses indicate that DWAT alleviates the out-of-bound degradation in the settings that we study, at a cost inside the training budget that we report and discuss. Full article

The end-winding insulation structure of stator windings in pumped-storage generator units is complex, with pronounced electric field concentration under out-of-phase conditions, making them critical concerns in insulation design and condition-based maintenance. Although the finite element method (FEM) offers reliable accuracy, the strong nonlinearity of the anti-corona layer results in a computation time exceeding 10⁴ seconds per single solution, rendering it impractical for parameter optimization and rapid on-site assessment. This paper proposes a fast prediction method for end-region potential distribution based on a deep neural network (DNN). Taking a 334 MW unit as the research object, a three-dimensional electroquasistatic finite element model with six stator coils is established and validated through power-frequency withstand voltage and ultraviolet imaging experiments. Training samples are generated via design of experiments (DoE), and a multilayer DNN surrogate model with a 7-dimensional input (comprising 3D spatial coordinates and four physical parameters) and a 1-dimensional output is constructed to directly reconstruct the spatial potential field at the end region. The results demonstrate that the surrogate model achieves a maximum relative error of less than 2% along the entire path compared with the high-fidelity FEM solutions, with a single prediction time of approximately 38 s—representing a speedup factor of approximately 272—while also exhibiting good generalization capability. This method provides a feasible technical approach for rapid reconstruction of end-region field distribution and optimization of insulation structures. Full article

As the crisis of antibiotic resistance escalates, phage therapy has regained attention as an alternative strategy. Artificial intelligence (AI) technologies offer new avenues to overcome the bottlenecks inherent in traditional bacteriophage research. This review summarizes the multi-dimensional innovative applications of machine learning, deep learning, and large biological models in phage studies. In the fields of phage recognition and genomics, support vector machines (SVMs), convolutional neural networks (CNNs), and pre-trained protein language models can all achieve recognition accuracy rates of over 90%. Furthermore, tools such as DeepHost and VirSorter2 can efficiently identify phage sequences, annotate functional genes, and predict hosts at the species or strain levels. For clinical translation, AI integrates patient characteristics, bacterial phenotypes, and phage profiles to customize cocktail regimens for individualized phage therapy. Graph neural network-based models like DeepPBI-KG integrate multi-omics knowledge graphs to precisely predict phage-host interactions (PHIs), whereas agent-based simulation and defense protein predictors forecast phage resistance evolution. Additionally, generative AI can support the de novo design of functional phage genomes and mine massive unannotated virome dark matter. Nevertheless, this cross-disciplinary field faces significant constraints, including uneven and biased sequencing datasets, insufficient model interpretability, and dual-use biosafety ethical risks accompanied by unclear algorithm accountability and incomplete global supervision systems. Future research should optimize standardized multimodal databases, develop explainable AI algorithms, and establish cross-disciplinary ethical governance frameworks to facilitate closed-loop verification between computational prediction and wet-lab experiments. In conclusion, the deep integration of AI and phage biology provides revolutionary strategies to tackle multidrug-resistant infections and advances the clinical transformation of phage precision medicine. Full article

Background/Objectives: Lateral flow assays (LFAs) are widely used point-of-care diagnostic devices due to their low cost, portability, and ease of use. However, most LFAs provide only qualitative results, limiting their utility for applications requiring continuous biomarker monitoring. This study introduces DeepBand, a deep learning-enabled multi-stage framework designed to automate the continuous quantification of analyte concentrations from unstandardized smartphone-captured lateral flow assay (LFA) images. Methods: A publicly available dataset containing 672 COVID-19 LFA images corresponding to four analyte concentrations (0.0, 1.8, 3.7, and 7.4 ng) was analyzed. A multi-stage pipeline was developed consisting of: (1) YOLOv11-based object detection to isolate the LFA cartridge from background artifacts, (2) a custom computer vision algorithm to identify and crop the test and control bands, and (3) a custom convolutional neural network (CNN) trained as a supervised regression model to predict continuous analyte concentrations. Data augmentation, hyperparameter optimization, and 5-fold cross-validation were used to improve model robustness. Results: The YOLOv11 model achieved approximately 99% mAP50 and 93.96% mAP95 for cartridge detection. Initial CNN models exhibited systematic underprediction of higher concentrations due to target imbalance; replacing mean squared error with Huber loss substantially improved performance, resulting in a final 20% held-out test set RMSE of 0.0292 ng. Analysis of HSV image channels demonstrated that the saturation-channel test-to-control intensity ratio was strongly correlated with analyte concentration (r = 0.94), consistent with the Beer–Lambert law governing LFA signal formation. Channel ablation studies confirmed the saturation channel as the most informative feature, while saliency mapping showed that the model primarily focused on biologically relevant test and control line regions. Conclusions: The proposed deep learning-enabled workflow, DeepBand, successfully integrates object detection, image processing, and CNN-based regression to provide automated quantitative interpretation of LFA results from smartphone images. Furthermore, the observed agreement between model behavior and Beer–Lambert theory suggests that the network learns biologically meaningful signal characteristics, supporting its potential for quantitative point-of-care diagnostics and longitudinal disease monitoring. Full article

Red mud, an alkaline industrial by-product of alumina refining generated in enormous volumes worldwide, poses a persistent environmental disposal challenge; using it as a partial cement replacement offers a promising route toward more sustainable concrete, but the resulting compressive strength is governed by complex, nonlinear interactions among the mix constituents that conventional empirical and regression-based models struggle to capture accurately. To address this challenge, the present study develops and compares four machine learning and deep learning models, namely the Deep Gradient Boosting Machine (DGBM), the Differentiable Neural Decision Tree (DNDT), Long Short-Term Memory (LSTM), and the Monte Carlo Dropout Neural Network (MCDNN), for the accurate and uncertainty-aware prediction of the compressive strength of red mud concrete. A dataset of 183 data points, compiled from the literature and supplemented with experimental results, was used to capture the influence of red mud content, curing period, and other mix parameters, including cement dosage, water content, and admixture proportions. The data were pre-processed prior to model training, and predictive performance was evaluated using R², RMSE, MAE, and WMAPE, among other indicators. The results show that the deep learning models outperformed the tree-based models: LSTM achieved the highest accuracy (R² = 0.942 on the testing dataset), while MCDNN additionally provided reliable uncertainty estimates alongside comparable prediction accuracy; DNDT and DGBM were comparatively less effective. Global sensitivity analysis identified fly ash and water content as the most influential contributors to strength development. By combining rigorous data-driven modeling with sensitivity and uncertainty analysis, this study contributes to the literature a validated, uncertainty-aware deep learning framework for sustainable concrete strength prediction, and offers practical value to the construction industry by providing engineers with a reliable, data-driven tool for optimizing red mud content in concrete mix design, thereby supporting the safe and wider industrial utilization of this problematic waste stream. Full article