Enhanced epileptic seizure detection using CNNs with convolutional block attention and short-term memory networks.

Characterization of Gastrodiae Rhizoma from different geographical origins by HS-GC-IMS and authenticity identification combined with deep learning.

Predicting ground motion amplification factors (AFs) is essential for constructing resilient infrastructure and managing earthquake risks in countries like India, which has a diverse range of geological and geophysical characteristics. The focus of this study was on predicting AFs across different soil sites and recorded ground motions using various machine learning (ML) methods. The training and testing datasets for AFs were created numerically using DeepSoil. A one-dimensional ground response analysis using equivalent linear methods was conducted at 124 soil sites across 20 regions of India. For this analysis, spectrum-compatible time histories were used. To predict AFs using ML models, it is vital to incorporate diverse parameters that represent both soil and seismic ground motion traits. Key indicators include the soil depth, shear wave velocity and fundamental period, while parameters such as magnitude, epicentral distance and peak ground acceleration illustrate ground motion features. In this research, an artificial neural network (ANN), random forest regression (RFR) and extreme gradient boosting (XGBoost) were trained and assessed for their predictive efficacy. It was found that the soil site's primary time period and shear wave velocity are crucial for predicting AFs. The ML models were found to excel at estimating AFs in Indian regions, with the XGBoost model outperforming the ANN and RFR models.

Deep learning techniques rely on artificial neural networks as their building blocks. This paradigm highlights the importance of neural networks for building models to address complex systems, including protein systems. We have successfully used neural networks to construct regression models to predict binding affinity based on atomic coordinates of protein-ligand complexes. Here, we focus on a neural network model to calculate the inhibition of cyclin-dependent kinase 2 (CDK2). This enzyme is a target for the development of anticancer drugs. To build our model, we employed the atomic coordinates of a CDK2-Cyclin A2 complex and the binding affinity data available at the BindingDB. We used the program Molegro Data Modeller to construct our regression model. Our model utilizes features determined by the Molegro Virtual Docker (MVD) program and shows superior predictive performance compared with classical scoring functions. All CDK2 datasets and Jupyter Notebooks discussed in this work are available at GitHub: https://github.com/azevedolab/docking#readme .

Deep Learning Approaches for Thrombosis Detection and Risk Assessment Via Ultrasound Imaging: A Scoping Review.

Memory-driven modeling of herpes simplex virus type-1 and type-2 dynamics with neural network optimization.

Customer-oriented knowledge management is a comprehensive approach aimed at developing a broad and integrated organizational vision, with its primary focus on achieving innovation and organizational effectiveness. This study examined customer-oriented knowledge management in technology-based companies located in Tehran using an artificial neural network approach. The research method was quantitative, survey-based, and applied in nature. Data were collected through a questionnaire administered to 386 managers and experts. To predict and evaluate patterns, a Multilayer Perceptron (MLP) neural network was utilized. The results indicated that input components such as customer-oriented knowledge management processes and behavioral data had strong correlations with output variables including customer satisfaction, innovation, and customer loyalty. The model demonstrated high predictive accuracy based on evaluation metrics such as Mean Absolute Error (MAE), Mean Squared Error (MSE), and the coefficient of determination (R²). The R² value of 0.83 reflected the model’s desirable performance. In the learning curve analysis, both training and testing errors decreased rapidly and stabilized, indicating optimal learning of the model and prevention of overfitting. The findings suggest that neural networks can serve as an effective tool for implementing customer-oriented knowledge management in technology-based companies, contributing to improved strategic decision-making processes and enhanced customer satisfaction.

The combination of machine learning (ML) and neural networks (NN), specifically deep learning (DL), is making a big breakthrough to breast cancer studies. This scientometrics survey studies 200 highly cited publications to map the intellectual landscape and studies trends in this dynamic field. The survey discloses a dominant concentration on computer-aided diagnosis (CAD) systems using convolutional neural networks (CNNs) for the classification of breast cancer from different imaging modalities, including mammography, histopathology, ultrasound, and magnetic resonance imaging (MRI). Key survey directions identified include: (1) the development of comprehensive deep learning techniques for image-based detection and classification; (2) the application of transfer learning to resolve data scarcity; (3) the combination of multi-omics and clinical data for personalized prognosis and treatment prediction; and (4) the exploration of explainability and robustness in ML-driven clinical tools. This study synthesizes the methodological advancements, sheds light on the evolution from traditional machine learning to deep learning, and surveys the challenges associated with data heterogeneity, model interpretability, and clinical integration. By giving a structured overview of the seminal work and emerging paradigms, the study serves as a reference for graduate students and other interested parties to have a better understanding about the current state and future trajectories of AI in breast oncology.

On the application of neural networks for structured domains to fMRI data.

Deep Fuzzy-NN modeling for the prediction of Zn(II) adsorption in columns using alkaline modified biochar: Integrated experimental and computational insights.

This study applies machine learning, specifically transfer learning with neural networks, to improve predictions of fission barrier heights and ground state binding energies of superheavy nuclei, which are crucial for calculating survival probabilities in fusion reactions. Transfer learning for neural networks proceeds in two stages: pre-training and fine-tuning, each driven by a distinct pre-training data set and target data set. In this work we split the pre-training data into 60 % for training and 40 % for validation, while the target data are partitioned into 20 % test, with the remaining 80 % further divided into 60 % training and 40 % validation. To construct the neural-network model we adopt the proton number Z and mass number A as the input layer, employ two hidden layers each containing 128 neurons with ReLU (Rectified Linear Unit) activation, and set the learning rate to 0.001. For the fission-barrier-height model, the pre-training dataset is either the FRLDM or the WS4 model data, and the experimental measurements serve as the target set. For the ground-state binding-energy model, we first form the residuals between WS4 predictions and the AME2020 evaluation, then separate these residuals into a light-nucleus subset and a heavy-nucleus subset according to proton number. The light-nucleus subset is used for pre-training and the heavy-nucleus subset for fine-tuning. After optimization, the root-mean-square error (RMSE) of the FRLDM barrier model falls from 1.03 MeV to 0.60 MeV, and that of the WS4 barrier model drops from 0.97 MeV to 0.61 MeV. For the binding-energy model, the RMSE decreases from 0.33 MeV to 0.17 MeV on the test set and from 0.29 MeV to 0.26 MeV on the full data set. We also present the performance of the fission-barrier model before and after refinement, together with the predicted barrier heights along the isotopic chains of the new elements Z = 119 and Z = 120, and analyzed the reasons for the differences in the results obtained by different models. We hope that these results are intended to provide a useful reference for future theoretical studies. The datasets in this paper are openly available at https://www.doi.org/10.57760/sciencedb.28388(Please use private access link https://www.scidb.cn/s/6fmeIz to access the dataset during the peer review process).

Recent advances in crosstalk simulation using integer-order memristive synapses have shown considerable progress. However, most existing models still employ a single-memristor structure, which constrains synaptic weight modulation and makes it difficult to represent both excitatory and inhibitory synaptic connections in a unified manner. These models also often fail to capture the memory effects and nonlocal dynamic properties inherent in biological neurons. To address these issues, this study introduces a fractional-order memristive bridge synapse model for crosstalk coupling. By combining Hindmarsh–Rose (HR) and FitzHugh–Nagumo (FN) neurons, we construct an 8D heterogeneous coupled neural network based on fractional calculus—designated as the Fractional-Order Memristive Bridge Crosstalk-Coupled Neural Network (FMBCCNN). A major innovation is the incorporation of a fractional-order memristive bridge structure that mimics synaptic connections in a bridge configuration. This design provides both historical memory characteristics and bidirectional synaptic weight regulation, overcoming limitations of traditional coupling forms.<br>Using dynamical analysis tools such as phase portraits, bifurcation diagrams, and Lyapunov exponents, we systematically investigate how synaptic and crosstalk strengths influence system behavior under conventional fractional-order conditions. The results reveal diverse dynamical behaviors, including attractor coexistence, forward and reverse period-doubling bifurcations, and chaotic crises. Further analysis under the more generalized condition of non-uniform fractional orders shows that, compared with the conventional case, the system maintains continuous periodic motion over broader parameter ranges and exhibits clear parameter hysteresis. Although local dynamic patterns remain similar, the corresponding parameter intervals are substantially widened. In addition, the system displays more concentrated and marked alternation between periodic and chaotic behaviors. We also simulate the effect of varying the fractional-order derivative, offering a more general mathematical characterization of neuronal firing activity.<br>Finally, the chaotic sequences generated by the system are applied to an image encryption algorithm incorporating bit-plane decomposition and DNA encoding. Security analysis confirms that the encrypted images have pixel correlation coefficients below 0.01 in horizontal, vertical, and diagonal directions, information entropy greater than 7.999, and a key space of 2<sup>2080</sup>. These results verify the excellent encryption performance and reliability of the proposed scheme and the generated sequences.

What makes good exemplars of a scene category good? Evidence from deep neural nets.

Blanket remote maintenance robot motion control based on nonlinear model predictive control and neural network

Dual-modal fusion of hierarchical image features and spectral data for efficient quantitative analysis of mineral elements in rice (Oryza sativa L.) leaves.

NEST: Neural estimation by sequential testing.

Sentiment analysis, a key component of natural language processing and data mining, plays a pivotal role in extracting subjective insights from textual data, particularly on social media platforms. In response to the growing importance of digital engagement, understanding public sentiment has become essential for formulating effective marketing strategies. This study aims to enhance the marketing strategy of energy products in subsidiaries of State-Owned Enterprises (SOEs) by employing a hybrid sentiment analysis model that integrates the Valence Aware Dictionary and Sentiment Reasoner (VADER) with Long Short-Term Memory (LSTM) neural networks. Utilizing a mixed-method approach that combines both quantitative and qualitative analyses, the study collects and processes data from multiple social media sources to identify and classify consumer sentiment. The results demonstrate that the hybrid VADER-LSTM model achieves an accuracy rate of up to 84%, enabling a more nuanced interpretation of consumer opinions. These insights inform the development of data-driven, responsive, and targeted marketing strategies. Furthermore, the study highlights the significance of fostering interactive communication between companies and consumers to enhance the impact of digital marketing efforts. Theoretical implications include a contribution to the academic discourse on information systems and digital marketing, while practical outcomes offer valuable guidance for SOEs in adopting adaptive, sentiment-informed marketing approaches within the energy sector.

Advanced modelling of green roof-enabled Plus-ZEB energy systems: A synergistic approach using Taguchi design and neural networks

Simplifying the operation of hybrid ad-hoc sensor networks with neural networks as the sole data reconstruction method

The rapid expansion of digital medical imaging technologies demands advanced tools for efficient and accurate image analysis. This research introduces a novel approach to medical image captioning, integrating convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to enhance the automatic generation of descriptive text for medical images. Our proposed model exploits the robust feature extraction capabilities of CNNs alongside the advanced sequential data processing of RNNs. We incorporate an attention mechanism that selectively focuses on diagnostically significant areas within images, thereby improving the relevance and accuracy of the generated captions. The effectiveness of our model was validated using an extensive set of evaluation metrics, including BLEU scores for linguistic quality and traditional classification metrics for accuracy. Results indicate that our model significantly outperforms existing systems in syntactic coherence and semantic accuracy, making it a valuable tool for aiding clinical decision-making and enhancing medical documentation.