Search Results (57)

Out-of-distribution (OOD) detection is essential for ensuring the reliability of machine learning models deployed in safety-critical applications. Existing methods often rely solely on statistical properties of feature distributions while ignoring the geometric structure of learned representations. We propose TopoAD, a topology-aware OOD detection framework that leverages Euler Characteristic Curves (ECCs) extracted from intermediate convolutional activation maps and fuses them with standardized energy scores. Specifically, we employ a computationally efficient superlevel-set filtration with a local estimator to capture topological invariants, avoiding the high cost of persistent homology. Furthermore, we introduce task-adaptive aggregation strategies to effectively integrate multi-scale topological features based on the complexity of distribution shifts. We evaluate our method on CIFAR-10 against four diverse OOD benchmarks spanning far-OOD (Textures), near-OOD (SVHN), and semantic shift scenarios. Our results demonstrate that TopoAD-Gated achieves superior performance on far-OOD data with 89.98% AUROC on Textures, while the ultra-lightweight TopoAD-Linear provides an efficient alternative for near-OOD detection. Comprehensive ablation studies reveal that cross-layer gating effectively captures multi-scale topological shifts, while threshold-wise attention provides limited benefit and can degrade far-OOD performance. Our analysis demonstrates that topological features are particularly effective for detecting OOD samples with distinct structural characteristics, highlighting TopoAD’s potential as a sustainable solution for resource-constrained applications in texture analysis, medical imaging, and remote sensing. Full article

Modeling flow field evolution accurately is important for numerous natural and engineering applications, such as pollutant dispersion in the ocean and atmosphere, yet remains challenging because of the highly nonlinear, multi-physics, and high-dimensional features of flow systems. While traditional equation-based numerical methods suffer from high computational costs, data-driven neural networks struggle with insufficient data and lack physical explainability. The physics-informed neural operator (PINO) addresses this by combining physics and data losses but faces a fundamental gradient imbalance problem. This work proposes a physics-informed fine-tuned neural operator for high-dimensional flow field modeling that decouples the optimization of physics and data losses. Our method first trains the neural network using data loss and then fine-tunes it with physics loss before inference, enabling the model to adapt to evaluation data while respecting physical constraints. This strategy requires no additional training data and can be applied to fit out-of-distribution (OOD) inputs faced during inference. We validate our method using the shallow water equation and advection–diffusion equation using a convolutional neural operator (CNO) as the base architecture. Experimental results show a 26.4% improvement in single-step prediction accuracy and a reduction in error accumulation for multi-step predictions. Full article

Background: Chest X-ray (CXR) is widely used for the assessment of thoracic diseases, yet automated multi-label interpretation remains challenging due to subtle visual patterns, overlapping anatomical structures, and frequent co-occurrence of abnormalities. While recent deep learning models have shown strong performance, limitations in interpretability, anatomical awareness, and robustness continue to hinder their clinical adoption. Methods: The proposed framework employs a hybrid ConvNeXtV2–Vision Transformer (ViT) architecture that combines convolutional feature extraction for capturing fine-grained local patterns with transformer-based global reasoning to model long-range contextual dependencies. The model is trained exclusively using image-level annotations. In addition to classification, three complementary post hoc components are integrated to enhance model trust and interpretability. A segmentation-aware Gradient-weighted class activation mapping (Grad-CAM) module leverages CheXmask lung and heart segmentations to highlight anatomically relevant regions and quantify predictive evidence inside and outside the lungs. An ontology-driven neuro-symbolic reasoning layer translates Grad-CAM activations into structured, rule-based explanations aligned with clinical concepts such as “basal effusion” and “enlarged cardiac silhouette”. Furthermore, a lightweight out-of-distribution (OOD) detection module based on confidence scores, energy scores, and Mahalanobis distance scores is employed to identify inputs that deviate from the training distribution. Results: On the VinBigData test set, the model achieved a macro-AUROC of 0.9525 and a Micro AUROC of 0.9777 when trained solely with image-level annotations. External evaluation further demonstrated strong generalisation, yielding macro-AUROC scores of 0.9106 on NIH ChestXray14 and 0.8487 on CheXpert (frontal views). Both Grad-CAM visualisations and ontology-based reasoning remained coherent on unseen data, while the OOD module successfully flagged non-thoracic images. Conclusions: Overall, the proposed approach demonstrates that hybrid convolutional neural network (CNN)–vision transformer (ViT) architectures, combined with anatomy-aware explainability and symbolic reasoning, can support automated chest X-ray diagnosis in a manner that is accurate, transparent, and safety-aware. Full article

Ship maneuverability models are typically defined by three degrees of freedom: surge, sway, and yaw. However, patrol vessels operating in riverine environments often exhibit significant roll motion during course changes, necessitating the inclusion of this dynamic. This study develops interpretable machine learning models capable of predicting vessel behavior in four degrees of freedom (4-DoF): surge, sway, yaw, and roll. A dataset of 125 h of simulated maneuvers was employed, including 29 h of out-of-distribution (OOD) conditions to test model generalization. Four models were implemented and compared over a 15-step prediction horizon: linear regression, third-order polynomial regression, a state-space model obtained via the N4SID algorithm, and an AutoRegressive model with eXogenous inputs (ARX). Results demonstrate that all models captured the essential vessel dynamics, with the state-space model achieving the best overall performance (e.g., NMSE = 0.0246 for surge velocity on test data and 0.0499 under OOD conditions). Variable-wise, surge and sway showed the lowest errors, roll rate remained stable, and yaw rate was the most sensitive to distribution shifts. Model-wise, the ARX model achieved the lowest NMSE for surge prediction (0.0149), while regression-based models provided interpretable yet less accurate alternatives. Multi-horizon evaluation (1-, 5-, 15-, and 30-step) under OOD conditions confirmed a consistent monotonic degradation across models. These findings validate the feasibility of using interpretable machine learning models for predictive control, autonomous navigation, and combat scenario simulation in riverine operations. Full article

Rice diseases pose a critical threat to global food security. While deep learning offers a promising path toward automated diagnosis, clear guidelines for model selection in resource-constrained agricultural environments are still lacking. This study presents a systematic benchmark of five deep convolutional neural networks (CNNs)—Visual Geometry Group (VGG)16, VGG19, Residual Network (ResNet)101V2, Xception, and Densely Connected Convolutional Network (DenseNet)121—for rice disease identification using a public leaf image dataset. The models, initialized with ImageNet pre-trained weights, were rigorously evaluated under a unified framework, including 5-fold cross-validation and a challenging out-of-distribution (OOD) generalization test. Our results demonstrate a clear performance hierarchy, with DenseNet121 emerging as the superior model. It achieved the highest OOD accuracy and F1-score (both 85.08%) while exhibiting the greatest parameter efficiency (8.1 million parameters), making it ideally suited for edge deployment. In contrast, architectures with large fully connected layers (VGG) or less efficient feature learning mechanisms (Xception, ResNet101V2) showed lower performance in this specific task. This study confirms the critical impact of architectural design choices, provides a reproducible performance baseline, and identifies DenseNet121 as a robust, efficient, and highly recommendable CNN for practical rice disease diagnosis in precision agriculture. Full article

Despite the success of reinforcement learning (RL), the common assumption of online interaction prevents its widespread adoption. Offline RL has emerged as an alternative that learns a policy from precollected data. However, this learning paradigm introduces a new challenge called “distributional shift”, degrading the performance of the policy when evaluated on out-of-distribution (OOD) scenarios (i.e., outside of the training data). Most existing works resolve this by policy regularization to optimize a policy within the support of the data. However, this overlooks the potential for high-reward regions outside of the data. This motivates offline policy optimization that is capable of finding high-reward regions outside of the data. In this paper, we devise a causality-based model architecture to accurately capture the OOD scenarios wherein the policy can be optimized without performance degradation. Specifically, we adapt causal normalizing flows (CNFs) to learn the transition dynamics and reward function for data generation and augmentation in offline policy learning. Based on the physics-based qualitative causal graph and precollected data, we develop a model-based offline OOD-adapting causal RL (MOOD-CRL) algorithm to learn the quantitative structural causal model. Consequently, MOOD-CRL can exercise counterfactual reasoning for sequential decision-making, revealing a high potential for OOD adaptation. The effectiveness is validated through extensive empirical evaluations with ablations including data quality and algorithmic sensitivity. Our results show that MOOD-CRL achieves comparable results with its online counterparts and consistently outperforms state-of-the-art model-free and model-based baselines by a significant margin. Full article

Valve internal leakage in thermal power stations exhibits a strong concealed nature. If it cannot be discovered and predicted of development trend in time, it will affect the safe and economical operation of plant equipment. This paper proposed an intelligent identification method for valve internal leakage that integrated an Improved Kepler Optimization Algorithm (IKOA) with Support Vector Regression (SVR). The Kepler Optimization Algorithm (KOA) was improved using the Sobol sequence and an adaptive Gaussian mutation strategy to achieve self-optimization of the key parameters in the SVR model. A multi-step sliding cross-validation method was employed to train the model, ultimately yielding the IKOA-SVR intelligent identification model for valve internal leakage quantification. Taking the main steam drain pipe valve as an example, a simulation case validation was carried out. The calculation example used Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and determination coefficient (R²) as performance evaluation metrics, and compared and analyzed the training and testing dataset using IKOA-SVR, KOA-SVR, Particle Swarm Optimization (PSO)-SVR, Random Search (RS)-SVR, Grid Search (GS)-SVR, and Bayesian Optimization (BO)-SVR methods, respectively. For the testing dataset, the MSE of IKOA-SVR is 0.65, RMSE is 0.81, MAE is 0.49, and MAPE is 0.0043, with the smallest values among the six methods. The R² of IKOA-SVR is 0.9998, with the largest value among the six methods. It indicated that IKOA-SVR can effectively solve problems such as getting stuck in local optima and overfitting during the optimization process. An Out-Of-Distribution (OOD) test was conducted for two scenarios: noise injection and Region-Holdout. The identification performance of all six methods decreased, with IKOA-SVR showing the smallest performance decline. The results show that IKOA-SVR has the strongest generalization ability and robustness, the best effect in improving fitting ability, the smallest identification error, the highest identification accuracy, and results closer to the actual value. The method presented in this paper provides an effective approach to solve the problem of intelligent identification of valve internal leakage in thermal power station. Full article

To address the problems of local optima and insufficient convergence accuracy in parameter identification of primary frequency regulation (PFR) for steam turbines, this paper proposed a hybrid identification method that integrated an Improved Bayesian Optimization (IBO) algorithm and an Improved Whale Optimization Algorithm (IWOA). By initializing the Bayesian parameter population using Tent chaotic mapping and the reverse learning strategy, employing a radial basis kernel function hyperparameter training mechanism based on the Adam optimizer and optimizing the Expected Improvement (EI) function using the Limited-memory Broyden–Fletcher– Goldfarb–Shanno with Bounds (L-BFGS-B) method, IBO was proposed to obtain the optimal candidate set with the smallest objective function value. By introducing a nonlinear convergence factor and the adaptive Levy flight perturbation strategy, IWOA was proposed to obtain locally optimized optimal solutions. By using the reverse-guided optimization mechanism and employing a fitness-oriented selection strategy, the optimal solution was chosen to complete the closed-loop process of reverse learning feedback. Nine standard test functions and the Proportional Integral Derivative (PID) parameter identification of the electro-hydraulic servo system in a 330 MW steam turbine were presented as examples. Compared with Particle Swarm Optimization (PSO), Whale Optimization Algorithm (WOA), Bayesian Optimization (BO) and Particle Swarm Optimization-Grey Wolf Optimizer (PSO-GWO), the Improved Bayesian Optimization-Whale Optimization Algorithm (IBO-WOA) proposed in this paper has been validated to effectively avoid the problem of getting stuck in local optima during complex optimization and has high parameter recognition accuracy. Meanwhile, an Out-Of-Distribution (OOD) Test based on noise injection had demonstrated that IBO-WOA had good robustness. The time constant identification of the steam turbine were carried out using IBO-WOA under two experimental conditions, and the identification results were input into the PFR model. The simulated power curve can track the experimental measured curve well, proving that the parameter identification results obtained by IBO-WOA have high accuracy and can be used for the modeling and response characteristic analysis of the steam turbine PFR. Full article

The Background/Objectives: The excessive dependence on data annotation, the lack of labeled data, and the substantial expense of data annotation, especially in healthcare, have constrained the efficacy of conventional supervised learning methodologies. Self-supervised learning (SSL) has arisen as a viable option by utilizing unlabeled data via pretext tasks. This paper examines the efficacy of supervised (pseudo-labels) and unsupervised (no pseudo-labels) pretext models in semi-supervised learning (SSL) for the classification of coronary artery disease (CAD) utilizing cardiac MRI data, highlighting performance in scenarios of data scarcity, out-of-distribution (OOD) conditions, and adversarial robustness. Methods: Two datasets, referred to as CAD Cardiac MRI and Ohio State Cardiac MRI Raw Data (OCMR), were utilized to establish three pretext tasks: (i) supervised Gaussian noise addition, (ii) supervised image rotation, and (iii) unsupervised generative reconstruction. These models were evaluated against  Simple Framework for Contrastive Learning (SimCLR), a prevalent unsupervised contrastive learning framework. Performance was assessed under three data reduction scenarios (20%, 50%, 70%), out-of-distribution situations, and adversarial attacks utilizing FGSM and PGD, alongside other significant evaluation criteria. Results: The Gaussian noise-based model attained the highest validation accuracy (up to 99.9%) across all data reduction scenarios and exhibited superiority over adversarial perturbations and all other employed measures. The rotation-based model exhibited considerable susceptibility to attacks and diminished accuracy with reduced data. The generative reconstruction model demonstrated moderate efficacy with minimal performance decline. SimCLR exhibited strong performance under standard conditions but shown inferior robustness relative to the Gaussian noise model. Conclusions: Meticulously crafted self-supervised pretext tasks exhibit potential in cardiac MRI classification, showcasing dependable performance and generalizability despite little data. These initial findings underscore SSL’s capacity to create reliable models for safety-critical healthcare applications and encourage more validation across varied datasets and clinical environments. Full article

In image classification tasks, models may occasionally produce incorrect predictions, which can lead to severe consequences in safety-critical applications. For instance, if a model mistakenly classifies a red traffic light as green, it could result in a traffic accident. Therefore, it is essential to assess the confidence level associated with each prediction. Predictions accompanied by high confidence scores are generally more reliable and can serve as a basis for informed decision-making. To address this, the present paper extends the block-scaling approach—originally developed for estimating classifier accuracy on unlabeled datasets—to compute confidence scores for individual samples in image classification. The proposed method, termed block masking confidence (BMC), applies a sliding mask filled with random noise to occlude localized regions of the input image. Each masked variant is classified, and predictions are aggregated across all variants. The final class is selected via majority voting, and a confidence score is derived based on prediction consistency. To evaluate the effectiveness of BMC, we conducted experiments comparing it against Monte Carlo (MC) dropout and a vanilla baseline across image datasets of varying sizes and distortion levels. While BMC does not consistently outperform the baselines under standard (in-distribution) conditions, it shows clear advantages on distorted and out-of-distribution (OOD) samples. Specifically, on the level-3 distorted iNaturalist 2018 dataset, BMC achieves a median expected calibration error (ECE) of 0.135, compared to 0.345 for MC dropout and 0.264 for the vanilla approach. On the level-3 distorted Places365 dataset, BMC yields an ECE of 0.173, outperforming MC dropout (0.290) and vanilla (0.201). For OOD samples in Places365, BMC achieves a peak entropy of 1.43, higher than the 1.06 observed for both MC dropout and vanilla. Furthermore, combining BMC with MC dropout leads to additional improvements. On distorted Places365, the median ECE is reduced to 0.151, and the peak entropy for OOD samples increases to 1.73. Overall, the proposed BMC method offers a promising framework for uncertainty quantification in image classification, particularly under challenging or distribution-shifted conditions. Full article

This paper investigates federated learning (FL) for cross-site scripting (XSS) detection under out-of-distribution (OOD) drift. Real-world XSS traffic involves fragmented attacks, heterogeneous benign inputs, and client imbalance, which erode conventional detectors. To simulate this, we construct two structurally divergent datasets: one with obfuscated, mixed-structure samples and another with syntactically regular examples, inducing structural OOD in both classes. We evaluate GloVe, GraphCodeBERT, and CodeT5 in both centralised and federated settings, tracking embedding drift and client variance. FL consistently improves OOD robustness by averaging decision boundaries from cleaner clients. Under FL scenarios, CodeT5 achieves the best aggregated performance (97.6% accuracy, 3.5% FPR), followed by GraphCodeBERT (96.8%, 4.7%), but is more stable on convergence. GloVe reaches a competitive final accuracy (96.2%) but exhibits a high instability across rounds, with a higher false positive rate (5.5%) and pronounced variance under FedProx. These results highlight the value and limits of structure-aware embeddings and support FL as a practical, privacy-preserving defence within OOD XSS scenarios. Full article

Domain shift poses a major challenge in many Machine Learning applications due to variations in data acquisition protocols, particularly in the medical field. Test-time augmentation (TTA) can solve the domain shift issue and improve robustness by aggregating predictions from multiple augmented versions of the same input. However, TTA may inadvertently generate unrealistic or Out-of-Distribution (OOD) samples that negatively affect prediction quality. In this work, we introduce a filtering procedure that removes from the TTA images all the OOD samples whose representations lie far from the training data distribution. Moreover, all the retained TTA images are weighted inversely to their distance from the training data. The final prediction is provided by a Self-Ensemble with Confidence, which is a lightweight ensemble strategy that fuses predictions from the original and retained TTA samples using a weighted soft voting scheme, without requiring multiple models or retraining. This method is model-agnostic and can be integrated with any deep learning architecture, making it broadly applicable across various domains. Experiments on cross-domain leukocyte classification benchmarks demonstrate that our method consistently improves over standard TTA and Baseline inference, particularly when strong domain shifts are present. Ablation studies and statistical tests confirm the effectiveness and significance of each component. Full article

The accurate estimation of building heights using very-high-resolution (VHR) synthetic aperture radar (SAR) imagery is crucial for various urban applications. This paper introduces a deep learning (DL)-based methodology for automated building height estimation from single VHR COSMO-SkyMed images: an object-based regression approach based on bounding box detection followed by height estimation. This model was trained and evaluated on a unique multi-continental dataset comprising eight geographically diverse cities across Europe, North and South America, and Asia, employing a cross-validation strategy to explicitly assess out-of-distribution (OOD) generalization. The results demonstrate highly promising performance, particularly on European cities where the model achieves a Mean Absolute Error (MAE) of approximately one building story (2.20 m in Munich), significantly outperforming recent state-of-the-art methods in similar OOD scenarios. Despite the increased variability observed when generalizing to cities in other continents, particularly in Asia with its distinct urban typologies and the prevalence of high-rise structures, this study underscores the significant potential of DL for robust cross-city and cross-continental transfer learning in building height estimation from single VHR SAR data. Full article

Although SAR systems can provide high-resolution aerial view images all-day, all-weather, the aspect and pose-sensitivity of the SAR target signatures, which defies the Gestalt perceptual principles, sets a frustrating performance upper bound for SAR Automatic Target Recognition (ATR). Therefore, we propose a network to support context-guided ATR by using aligned Electro-Optical (EO)-SAR image pairs. To realize EO-SAR image scene grammar alignment, the stable context features highly correlated to the parking patterns of the vehicle and aircraft targets are extracted from the EO images as prior knowledge, which is used to assist SAR-ATR. The proposed network consists of a Scene Recognition Module (SRM) and an instance-level Cross-modality ATR Module (CATRM). The SRM is based on a novel light-condition-driven adaptive EO-SAR decision weighting scheme, and the Outlier Exposure (OE) approach is employed for SRM training to realize Out-of-Distribution (OOD) scene detection. Once the scene depicted in the cut of interest is identified with the SRM, the image cut is sent to the CATRM for ATR. Considering that the EO-SAR images acquired from diverse observation angles often feature unbalanced quality, a novel class-incremental learning method based on the Context-Guided Re-Identification (ReID)-based Key-view (CGRID-Key) exemplar selection strategy is devised so that the network is capable of continuous learning in the open-world deployment environment. Vehicle ATR experimental results based on the UNICORN dataset, which consists of 360-degree EO-SAR images of an army base, show that the CGRID-Key exemplar strategy offers a classification accuracy 29.3% higher than the baseline model for the incremental vehicle category, SUV. Moreover, aircraft ATR experimental results based on the aligned EO-SAR images collected over several representative airports and the Arizona aircraft boneyard show that the proposed network achieves an F1 score of 0.987, which is 9% higher than YOLOv8. Full article

Detecting stored-grain pests on the surface of the grain pile plays an important role in integrated pest management (IPM), which is crucial for grain security. Recently, numerous deep learning-based pest detection methods have been proposed. However, a critical limitation of existing methods is their inability to detect out-of-distribution (OOD) categories that are unseen during training. When encountering such objects, these methods often misclassify them as in-distribution (ID) categories. To address this challenge, we propose a one-stage framework named PestOOD for out-of-distribution stored-grain pest detection via flow-based feature reconstruction. Specifically, we propose a novel Flow-Based OOD Feature Generation (FOFG) module that generates OOD features for detector training via feature reconstruction. This helps the detector learn to recognize OOD objects more effectively. Additionally, to prevent network overfitting that may lead to an excessive focus on ID feature extraction, we propose a Noisy DropBlock (NDB) module and integrate it into the backbone network. Finally, to ensure effective network convergence, a Stage-Wise Training Strategy (STS) is proposed. We conducted extensive experiments on our previously established multi-class stored-grain pest dataset. The results show that our proposed PestOOD demonstrates superior performance over state-of-the-art methods, providing an effective AI-enabled solution to ensure grain security. Full article