Artificial neural networks are reliable machine learning models that have been applied to a multitude of practical and scientific applications in recent decades. Among these applications, there are examples from the areas of physics, chemistry, medicine, etc. To effectively apply them to these problems, it is necessary to adapt their parameters using optimization techniques. However, in order to be effective, optimization techniques must know the range of values for the parameters of the artificial neural network, so that they can adequately train the artificial neural network. In most cases, this is not possible, as these ranges are also significantly affected by the inputs to the artificial neural network from the objective problem it is called upon to solve. This situation usually results in artificial neural networks becoming trapped in local minima of the error function or, even worse, in the phenomenon of overfitting, where although the training error achieves low values, the artificial neural network exhibits low performance in the corresponding test set. To address this limitation, this work proposes a novel two-stage training approach in which a simulated annealing (SA)-based preprocessing stage is employed to automatically identify optimal parameter value intervals before the application of any optimization method to train the neural network. Unlike similar approaches that rely on fixed or heuristically selected parameter bounds, the proposed preprocessing technique explores the parameter space probabilistically, guided by a temperature-controlled acceptance mechanism that balances global exploration and local refinement. The proposed method has been successfully applied to a wide range of classification and regression problems and comparative results are presented in detail in the present work.

Objectives . This study focuses on the development and investigation of a generalized nonlinear Support Vector Machine (SVM) method incorporating an adaptive transformation of the feature space. Its aim is to improve computational efficiency while maintaining high classification accuracy. The binary classification problem is used as a case study. The main objective of the research is to quantitatively evaluate the performance of the proposed approach when compared to classical SVM models using fixed kernel functions, and to analyze how the transformation parameters affect classification quality. Methods . The proposed approach involves a preliminary transformation of the input data using a learnable nonlinear mapping with a fixed structure. This mapping is implemented as a composition of elementary functions and is parameterized by a limited number of trainable weights which allows control over model complexity. A linear SVM with L2 regularization is applied after the transformation. The model is trained using conventional, unconstrained numerical optimization methods. The classification quality is evaluated using the Accuracy metric averaged over 10-fold cross-validation. The work also studies the behavior of the model with varying feature space dimensionality. In addition, computational complexity is analyzed in terms of the number of operations and inference time required on test datasets. Results . Numerical experiments demonstrate that the proposed model significantly reduces classification time when compared to a polynomial-kernel SVM, while maintaining a comparable level of accuracy. The runtime analysis confirms that the proposed approach scales much better than traditional kernel methods. At the same time, the structure of the model remains interpretable and can be further adapted to the specifics of the application domain. Conclusions . The method developed provides an efficient alternative to traditional kernel-based algorithms. Through the use of a parameterized transformation of the feature space, the method enables adaptability, interpretability, and scalability, making it promising for practical applications in machine learning tasks.

Evidential deep learning (EDL) models, based on Subjective Logic, introduce a principled and computationally efficient way to make deterministic neural networks uncertainty-aware. The resulting evidential models can quantify fine-grained uncertainty using learned evidence. However, the Subjective-Logic framework constrains evidence to be non-negative, requiring specific activation functions whose geometric properties can induce activation-dependent learning-freeze behavior-a regime where gradients become extremely small for samples mapped into low-evidence regions. We theoretically characterize this behavior and analyze how different evidential activations influence learning dynamics. Building on this analysis, we design a general family of activation functions and corresponding evidential regularizers that provide an alternative pathway for consistent evidence updates across activation regimes. Extensive experiments on four benchmark classification problems (MNIST, CIFAR-10, CIFAR-100, and Tiny-ImageNet), two few-shot classification problems, and blind face restoration problem empirically validate the developed theory and demonstrate the effectiveness of the proposed generalized regularized evidential models.

In this paper, we propose a new nonparallel support vector machine for binary classification problems and name it the improved nonparallel support vector machine (IMNSVM). The IMNSVM uses a one-sided ε-band and minimizes ε to achieve a better fitting effect for the same class of training points. By introducing a new variable, ρ, the IMNSVM keeps one class of training points at a certain distance from the hyperplane corresponding to another class of training points, keeping them as far away as possible so as to better adapt to the training points and better describe the difference in data distribution between different categories. The IMNSVM can degenerate into the standard support vector machine (SVM) under certain conditions and is applicable to a wider range of data types. Finally, numerical experiments also explain the effectiveness of the method.

Small sphere and large margin support tensor machines for imbalanced tensor data classification.

Regret theory-based clinical efficacy evaluation method with three-way decision.

Convergence analysis of recurrent neural networks based on sparse mechanism and its application to time series and multiple classification problems

Multiview-cooperated graph neural network enables novel multi-omics cancer subtype classification.

The development of artificial intelligence combined with microcontroller-based systems proves to be a breakthrough method in the precision farming. It allows smart decision-making and optimizes the usage of resources with real-time processing capability. This review describes the current system of AI-based systems in leaf age and maturity identification in most detail, with special focus on microcontroller inclusion, computing difficulties, and use cases. The review systematically examines many Artificial Intelligence methods such as convolutional Artificial Intelligence, machine learning algorithms, and edge computing strategies in solving leaf classification problems. Key advantages of microcontroller-based AI systems include cost-effectiveness 60–70% reduction compared to server-based systems low power consumption <50 mW, real-time decision-making capabilities >90% latency reduction, and cloud independence > 95% uptime in poor connectivity areas. Nevertheless, there are still serious issues such as the lack of computing resources, changes in the environment, and special skills. A combination of IoT, blockchain and enhanced sensor networks should provide the opportunities to improve the data security, the positioning the supply chain and deploy it at scale. The future research directions improve hybrid AI architectures, federated learning, and optimizing/training with a quantum advantage to address the computational bottleneck challenges favouring ecosustainability and efficiency in AI. The review presents a guideline to other researchers and practitioners as to how they can develop their own independent data-driven agricultural systems to be able to attain higher productivity, with reduced damage to the environment.

A method for a priori selection of the optimal Lee filter window size for speckle noise suppression in Sentinel-1 SAR images is presented. The task is formulated as a direct multiclass classification problem. The proposed model employs transfer learning with a compact MobileNetV2 backbone adapted to single-channel SAR input and equipped with a lightweight classification head. The training corpus is generated synthetically: Sentinel-2 patches are aligned to the Sentinel-1 distribution via histogram matching and subjected to dynamic multiplicative noise with gamma distribution and controlled spatial correlation. Ground truth labels are determined by exhaustive Lee filtering with window sizes of 3×3, 5×5, 7×7, 9×9, and 11×11, selecting argmax over the full quality metric. Six independent experiments were conducted for PSNR, SSIM, MS-SSIM, FSIM, HaarPSI, and MDSI criteria. On a test set of 382 patches (1024×1024), the best overall accuracy reached 87.17% (FSIM), balanced accuracy achieved 88.94% (MS-SSIM), minimum calibration error ECE was 0.0371 (PSNR), Brier score was 0.0359 (FSIM), with Top-2 accuracy in the range of 99.21–100.00% (100.00% for PSNR). Across ENL bins, accuracy remained and reached 91.1% for with PSNR and FSIM; across speckle spatial frequency scale, peaks of 93.7% (FSIM) at and 93.0% (MS-SSIM) at were observed. Critically, the method does not require a priori knowledge of the equivalent number of looks (ENL), which is consistent with accuracy stability in the range . The method is suitable for operational integration into production SAR pipelines for preprocessing and computational resource planning.

Background The use of ensemble learning has been crucial for improving predictive accuracy in healthcare, especially with regard to critical diagnostic and classification problems. Ensemble models combine the strengths of multiple ML models and reduce the risk of misclassification, which is important in healthcare, where accurate predictions impact patient outcomes. Methods This study introduces the Gradient-Based Weight Optimized Ensemble Model (GBWOEM), an advanced ensemble technique that optimizes the weights of five base models: Decision Tree Classifier (DTC), Random Forest Classifier (RFC), Logistic Regression (LR), Multi-Layer Perceptron (MLP), and K-Nearest Neighbours (KNN), through optimizing the weights. Two variants, GBWOEM-R (random weight initialization) and GBWOEM-U (uniform weight initialization), were proposed and tested on five healthcare-related datasets: breast cancer, Pima Indians Diabetes Database, diabetic retinopathy debrecen, obesity level estimation based on physical condition and eating habits, and thyroid diseases. Results The test accuracy of the proposed models increased to 0.48-8.26% over the traditional ensemble models, such as Adaboost, Catboost, GradientBoost, LightGBM, and XGBoost. Performance metrics, including ROC-AUC analyses, confirmed the model’s efficacy in handling imbalanced data, highlighting its potential for advancing predictive consistency in healthcare applications. Conclusion The GBWOEM model improves the predictive accuracy and offers a reliable solution for healthcare applications even when dealing with the imbalance data. This strategy has the potential to ensure patient outcomes and diagnostic consistency in healthcare settings.

Analysis of satellite images is an essential role in the response and recovery of disaster situations especially after a natural disaster like the hurricane. The conventional approaches of assessing damage heavily depend on manual examination and the antique paradigm of classification having the lexicon of limitations of being scalable, versatile and expedient. This paper presents the hybrid machine learning framework to focus on the problems of post-hurricane satellite images segmentation and classification. The solution being proposed is one which interpolates a Gradient-Hopping Hybrid Optimizer (GHHO) and a Pattern-Adaptive Classifier Fusion (PACF) mechanism. GHHO amalgamates, global exploration, and local exploitation as stochastic perturbation and local gradient descent, respectively, to maximize the optimization of segmentation parameters. At the same time, PACF will be adaptively choosing and fusing different classifiers with different feature subsets, which allows the model to react competently to the differences in land structures, atmospheric conditions, and sensor features. To improve the results further we propose a modification called Multi-Objective Gradient-Hopping Hybrid Optimizer (MO-GHHO) + PACF that involves multi-objective optimization to find improved convergence, generalization, and accuracy. The strength and versatility of both models is supported by experiments on high-resolution post-hurricane satellite datasets. The GHHO + PACF model reached the overall accuracy of 95.2% and surpassed the traditional architectures like Visual Geometry Group (VGG)-19, ResNet-50, Inception V3. Also, MO-GHHO + PACF achieves 98.2% of the land classification and 96.1% of the water body segmentation in addition to its significant increase in precision, recall, and F1 score. An ablation study also indicates the effectiveness in the individual contribution of the GHHO and PACF components in the model. The results indicate that the frameworks of GHHO + PACF and MO-GHHO + PACF have a robust, scalable, flexible application to the post-disaster analysis of satellite images and have a great potential as the decision support mechanism to any future rapid damage assessment systems.

Mango is a high-value agricultural commodity, and accurate and efficient appearance quality grading and defect inspection are critical for export-oriented markets. This study proposes a dual-pipeline deep learning framework for automated mango assessment, in which surface defect classification and quality grading are jointly implemented within a unified inspection system. For defect assessment, the task is formulated as a multi-label classification problem involving five surface defect categories, eliminating the need for costly bounding box annotations required by conventional object detection models. To address the severe class imbalance commonly encountered in agricultural datasets, a copy–paste-based image synthesis strategy is employed to augment scarce defect samples. For quality grading, mangoes are categorized into three quality levels. Unlike conventional CNN-based approaches relying solely on spatial-domain information, the proposed framework integrates decision-level fusion of spatial-domain and frequency-domain representations to enhance grading stability. In addition, image preprocessing is investigated, showing that adaptive contrast enhancement effectively emphasizes surface textures critical for quality discrimination. Experimental evaluations demonstrate that the proposed framework achieves superior performance in both defect classification and quality grading compared with existing detection-based approaches. The proposed classification-oriented system provides an efficient and practical integrated solution for automated mango assessment.

ABSTRACT This study proposes a theory‐informed learning analytics approach for predicting at‐risk students using large‐scale behavioural, demographic and academic data. Building on engagement theory and self‐regulated learning, we engineer pedagogically grounded behavioural indicators that move beyond raw click counts. These indicators include multidimensional engagement measures, temporal regularity and an antiprocrastination score derived from assessment submission patterns. Using the Open University Learning Analytics Dataset (OULAD), comprising 32,593 students across 22 courses, we reformulate the prediction task as a binary classification problem (Pass vs. At‐Risk) and compare three machine learning algorithms: Support Vector Machine (SVM), Random Forest (RF) and Extreme Gradient Boosting (XGBoost). Models are evaluated at four quarter‐based checkpoints over the semester to investigate temporal dynamics and opportunities for timely intervention. Results show that XGBoost consistently outperforms RF and SVM in accuracy, recall, precision and ROC AUC, while behavioural features overwhelmingly dominate demographic and academic variables in predictive importance. Temporal analysis reveals that model performance improves substantially from the first to the third quarter, with mid‐semester predictions offering the best trade‐off between accuracy and time remaining for effective support. The findings demonstrate the value of theory‐driven feature engineering and temporally sensitive evaluation in designing early‐warning systems that are both accurate and pedagogically actionable.

Nowadays, deep learning models are quickly increasing in complexity and size, making the training or inference phases unsuitable for most office computers or even high-performance computing servers. Examples of such models include, but are not limited to, large language models and image diffusion models, where the number of parameters to adjust can reach hundreds of billions. Although deep learning models are extremely useful tools, the need for powerful computers to train and use them might compromise the obtained results and also reduce accessibility for many researchers. In this work, we propose a general method to reduce the training time and computing resources (central processing unit and memory usage) by the division of a large-size model into several smaller submodels that are easier to handle but do not affect the final performance. This division depends on the original model and architecture, and hence we propose specific strategies for regression and classification problems. The main result of this work is the development of a public Python library called Skynnet to offer this alternative to deep learning computations.

BackgroundCurrently, there is a global mpox outbreak. It is imperative to develop predictive diagnostic models that incorporate clinical information to effectively manage outbreak.Figure 1.The ROC curve of lightGBM classifiers trained to predict PCR resultsLightGBM to predict mpox showed a sensitivity of 0.94, a precision of 0.88, a specificity of 0.56, and an area under the receiver operating characteristic curve of 0.75.MethodsThis prospective cohort study was conducted at the National Center for Global Health and Medicine, a national reference center for emerging infectious diseases in Japan, from July 2022 to July 2024. The study population included patients suspected of having mpox, based on their travel and contact histories or clinical symptoms. We analyzed the data as a binary classification problem using patient demographics, medical histories, and symptoms as explanatory variables, with the PCR test results of mpox (positive or negative) as the response variable. Using a light Gradient Boosting Machine (lightGBM) with leave-one-out cross-validation, we identified significant predictive factors using Power-Full Shapley Feature Selection Method (Powershap), repeated across 100 iterations with varying seeds. The relationship among patient characteristics between positive and negative patients in the PCR results was assessed using the Mann-Whitney U and Fisher's exact tests.Figure 2.Feature importance by Powershap selection frequency in leave-one-out CV using 100 seedsFeature selection using Powershap revealed that seven of the 105 features were sufficient for prediction. Skin itching and rash on the trunk at the first visit were thought to be important clinical features in the diagnosis of mpox.ResultsOf these 80 patients, 63 with mpox (clade IIb) and 17 without mpox were included (Table 1). All were male; 58 men who have sex with men were included in the mpox group and 15 in the non-mpox group, and only three mpox patients had CD4 counts < 200 cells/mm. Using lightGBM to predict mpox, we achieved a sensitivity of 0.94, a precision of 0.88, a specificity of 0.56, and an area under the receiver operating characteristic curve of 0.75 (Fig. 1). Feature selection using Powershap revealed that seven of the 105 features were sufficient for prediction (Fig. 2). Furthermore, when comparing factors between mpox and non-mpox patients, we found that skin itching and lymphadenopathy were significantly more prevalent in mpox patients, while non-mpox patients had a significant rash on the trunk at the first visit (Table 1).ConclusionAmong the clinically suspected cases of mild mpox, skin itching, lymphadenopathy, and rash on the trunk at the first visit were thought to be important clinical features in the diagnosis of mpox. Further research is needed to evaluate more appropriate predictive diagnostic models of mpox with large sample sizes and in hospitals.DisclosuresAll Authors: No reported disclosures

MT-IDS: A multi-task information decoupling strategy for identifying lymph node metastasis in the mediastinal region.

BackgroundFrontotemporal dementia (FTD) is frequently misdiagnosed as Alzheimer's disease (AD) due to their overlapping symptoms, leading to a decreased quality of life and misallocation of resources. Research indicates that neuroimaging techniques outperform cognitive tests in differentiating between these conditions, with electroencephalography (EEG) offering a cost‐effective, accessible, and faster alternative. Moreover, resting‐state EEG is less taxing on patients, which is particularly important for those with cognitive impairments.While EEG‐based classification between AD and healthy controls (HC), as well as FTD and HC, has shown promising results, accurately distinguishing AD from FTD remains challenging. Functional connectivity (FC) graphs are key features used to construct machine learning (ML) models for this problem. However, most of the research using ML with FC features has been conducted using flat Euclidean metrics, even though some connectivity graphs naturally reside on non‐Euclidean manifolds. Riemannian geometry provides a more suitable framework for analysing and classifying complex data structures, like some FC graphs, enhancing ML performance.MethodWe propose a classification strategy using Riemannian geometry and resting‐state EEG to differentiate between AD, FTD, and HC (Figure 1). This method utilizes stacked generalization to combine predictions obtained by a set of Riemannian geometry classifiers each trained on one of various FC features across different frequency bands (delta, theta, alpha, beta, gamma). A logistic regression meta‐classifier combined with sequential feature selection is used to reach a decision based on the predictions per FC feature.ResultOur proposed model was evaluated on an openly available dataset (doi:10.18112/openneuro.ds004504.v1.0.7) with 36 AD, 23 FTD, and 29 HC age‐matched subjects. Leave‐one‐subject‐out cross validation yielded the following classification performances for 3 binary classification problems while accounting for class imbalance: AD/FTD: 74%, AD/HC: 85%, FTD/HC: 74%.ConclusionBy using smaller, efficient models that account for data geometry and a meta‐classifier, our proposed method reaches good performance while allowing for enhancement through alternative feature selection and hyperparameter optimization. Additionally, our method identifies the most informative FC graphs for classification, which are valuable for model interpretation and can provide specific distinctive neural signals that can be used in future research to develop diagnostic tools.

Abstract Dense associative memory (DAM) models have been attracting renewed attention since they were shown to be robust to adversarial examples and closely related to cutting edge machine learning paradigms, such as the attention mechanism and generative diffusion. We study a DAM built upon a three-layer Boltzmann machine with Potts hidden units, which represent data clusters and classes. Through a statistical mechanics analysis, we derive saddle-point equations that characterize both the stationary points of DAMs trained on real data and the fixed points of DAMs trained on synthetic data within a teacher-student framework. Based on these results, we propose a novel regularization scheme that makes training significantly more stable. Moreover, we show empirically that our DAM learns interpretable solutions to both supervised and unsupervised classification problems. Pushing our theoretical analysis further, we find that the weights learned by relatively small DAMs correspond to unstable saddle points in larger DAMs. We implement a network-growing algorithm that leverages this saddle-point hierarchy to drastically reduce the computational cost of training dense associative memory.

DeepVariant has revolutionized the field of genetic variant identification by reframing variant detection as an image classification problem. However, despite its wide adoption in bioinformatics workflows, the tool continues to evolve mainly through the expansion of training datasets, while its core neural network architecture—Inception V3—has remained unchanged. In this study, we revisited the DeepVariant design and presented a prototype of a modernized version that supports alternative neural network backbones. As a proof of concept, we replaced the legacy Inception V3 model with a mid-sized EfficientNet model and evaluated its performance using the benchmark dataset from the Genome in a Bottle (GIAB) project. Alternative architecture demonstrated faster convergence, a twofold reduction in the number of parameters, and improved accuracy in variant identification. On the test dataset, updated workflow achieved consistent improvements of +0.1% in SNP F1-score, enabling the detection of up to several hundred additional true variants per genome. These results show that optimizing the neural architecture alone can enhance the accuracy, robustness, and efficiency of variant calling, thereby improving the overall quality of sequencing data analysis.