Shahnorbanun Sahran

The activation function is a critical component in convolutional neural networks that perform nonlinear mappings between inputs and outputs. The rectified linear unit (ReLU) function is currently the baseline activation function in the deep learning community; however, it has several drawbacks that result in inefficient training, such as how the zero value of a negative part treats negative inputs as insignificant information for learning, which results in bias shift in network layers, and how the multilinear structure of ReLU restricts the nonlinear approximation power of networks. To address these limitations, this study proposes the parametric arctangent unit (PATU), which integrates advantageous features of ReLU, (parametric ReLU) PReLU, (exponential linear unit) ELU and Swish. Unlike the PReLU, which employs a learnable linear slope for negative inputs and thus remains piecewise linear, PATU integrates a nonlinear arctangent activation in the negative part. Specifically, PATU retains ReLU’s advantageous positive part, while the negative part employs arctangent to activate negative inputs, and a trainable parameter is implemented to control the saturation of the negative part. The study compared the proposed PATU to four state-of-art activation functions, namely ReLU, PReLU, ELU and Swish, on the CIFAR10 and CIFAR100 datasets using SmallNet, Network in Network, and ResNet18 convolutional neural network architectures. Experimental results demonstrate that PATU outperformed ReLU in terms of accuracy on the CIFAR10 and CIFAR100 datasets by 0.45% and 1.44%, respectively, for the SmallNet architecture; 1.87% and 3.11%, respectively, for the NIN architecture; and 1.19% and 2.92%, respectively, for the ResNet18 architecture. All experiments were repeated five times, and the improvements were statistically significant (Friedman test, p = 0.015). PATU reached the highest mean rank score of 1.16 in the Friedman test for classification accuracy, which is superior to state-of-the-art methods. These results suggest that the proposed PATU is an effective alternative to ReLU.

In this paper, two novel and general predefined‐time stability (PTSt) lemmas are introduced and applied to address the predefined‐time synchronization (PTSy) problem in memristor‐based bidirectional associative memory (BAM) neural networks. Unlike traditional finite‐time synchronization (FinTS) or fixed‐time synchronization (FixTS) approaches, the synchronization time in this study is independent of the initial states (InSts) and system parameters. It can be predetermined according to task requirements without any estimation. Two effective and innovative controllers are designed to achieve synchronization in time‐delayed memristor‐based BAM neural networks (MBAMNNs) based on PTSt. The proposed controllers employ general PTSy analysis to ensure robust synchronization performance. These diverse predefined‐time controllers (PTCos) design facilitates the practical implementation of synchronization control in neural networks. By employing the fraction‐order power function and exponential function, the PTCos offer flexibility and adaptability in controller design. The results are validated through two numerical examples using MATLAB simulations.

The rising occurrence of mental health problems requires the creation of detection systems that are not only efficient but also understandable. This research utilizes the feature selection to enhance the performance of NLP and ML models for identifying mental health conditions. The suggested technique proposes conversational features combination that includes a hybrid mixture of traditional linguistic features (e.g., word frequency, and sentiment polarity) and novel and interactional features (e.g., conversational dynamics, response latency, and linguistic style matching) to get better results. These features are processed by advanced selection methods to ensure the interpretability and robustness of the model and integrated with XGboost to measure the effectiveness of the feature set and their impact. Various machine learning algorithms, including Support Vector Machines (SVM), Random Forest, Deep Neural Networks (DNNs), and XGBoost, were tested on publicly available text-based datasets to measure the effectiveness of the feature set and their impact. The proposed model experimental results show that the XGBoost model, which led to 91% accuracy, was the best performing model thus highlighting the importance of feature selection in facilitating NLP-based mental health detection. The proposed framework is an add on towards the creation of dependable and scalable AI-driven tools for early mental health assessment

Financial fraud has expanded rapidly with the growth of the digital economy, evolving from conventional transactional misconduct to more complex and data-intensive forms. Traditional rule-based detection methods are increasingly inadequate for addressing the scale, heterogeneity, and dynamic behavior of modern fraud. In this context, artificial intelligence (AI) has become a core tool in financial fraud detection research. This review systematically surveys AI-based financial fraud detection studies published between 2015 and 2025. It summarizes representative machine learning and deep learning approaches, including tree-based models, neural networks, and graph-based methods, and examines their applications in major fraud scenarios such as credit card fraud, loan fraud, and anti-money laundering. In addition, emerging research on cryptocurrency- and blockchain-related fraud is reviewed, highlighting the distinct challenges posed by decentralized transaction environments. Through a comparative analysis of methods, datasets, and evaluation practices, this review identifies persistent issues in the literature, including severe class imbalance, concept drift, limited access to labeled data, and trade-offs between detection performance and interpretability. Based on these findings, the paper discusses practical considerations for applied fraud detection systems and outlines future research directions from a data-centric and application-oriented perspective. This review aims to provide a structured reference for researchers and practitioners working on real-world financial fraud detection problems.

Data augmentation (DA) techniques are widely used in convolutional neural networks (CNNs) to artificially expand the size of training datasets. This is particularly true for medical imaging tasks, such as Diabetic Retinopathy (DR) image classification, where the training data are often limited and imbalanced. Various DA techniques are utilized in CNN models, including horizontal and vertical flipping, rotation, and zoom. Combining distinct methods increases the diversity of the produced images and allows the CNNs to handle the complex details in the images. Manually designed or heuristically selected augmentation combinations may generate redundant or highly similar training samples, which can negatively affect model generalization and classification performance. To analyze the effect of DA combinations on DR classification, the selection of the augmentation subset is formulated as a binary combinatorial optimization problem, in which the recent metaheuristic White Shark Optimizer (WSO) is employed to identify the optimal augmentation subset. Since the original WSO was designed for continuous optimization, the proposed Binary WSO (BWSO) adapts the WSO to binary search spaces through a sigmoid-based binary transfer mechanism. In each iteration of the WSO, candidate augmentation subsets are evaluated using the validation performance of three pre-trained CNN architectures (VGG16, DenseNet121, and MobileNetV2) to identify augmentation combinations that improve classification robustness and generalization. The proposed Adaptive Metaheuristic Data Augmentation–White Shark Optimizer (AMDA-WSO) framework was evaluated on the APTOS 2019 dataset containing 3662 retinal fundus images in five severity grades of DR and compared with Particle Swarm Optimization (PSO) and the Bees Algorithm (BA). Experimental results demonstrated that the AMDA-WSO-guided MobileNetV2 achieved the best performance, reaching 92.08% accuracy and 81.48% sensitivity. These findings confirm the effectiveness of BWSO for automated binary augmentation subset selection in DR image classification.

The optimization of hyperparameters in convolutional neural networks (CNNs) is crucial for improving the accuracy and effectiveness of diabetic retinopathy (DR) image classification. Conventional approaches often rely on manual adjustments, which are time-consuming and inefficient. This research proposes variant particle swarm optimization (PSO) methods for fine-tuning the hyperparameters of pre-trained deep learning models—specifically PSO-VGG-16, PSO-DenseNet-121, and PSO-MobileNet. The proposed models enhance the exploratory capabilities of CNNs by searching for optimal hyperparameter configurations. Experiments were conducted using the APTOS 2019 DR image benchmark dataset. This study introduces a PSO-driven hyperparameter optimization framework tailored for pre-trained CNN architectures (VGG-16, DenseNet-121, and MobileNetV2) to improve DR classification performance. The proposed method outperforms conventional genetic algorithm (GA)-based tuning techniques, achieving superior accuracy and model robustness. The findings show that PSO achieved a classification accuracy of 97.41%, sensitivity of 96.8%, and specificity of 99.4%. Among the models, PSO-MobileNet demonstrated the best overall performance. These results confirm that PSO is a powerful tool for hyperparameter optimization, enhancing diagnostic accuracy and improving the efficiency of CNN models in medical image classification. This technique supports healthcare professionals in making timely and precise diagnoses, contributing to better patient outcomes and reduced strain on healthcare systems.

The demand for robust unsupervised anomaly detection in streaming data has grown significantly in the era of smart devices, where vast amounts of data are continuously collected from such devices. Leveraging this data through effective anomaly detection is essential and necessitates a system that can work in real-time. One of the most innovative solutions is the Online Evolving Spiking Neural Network (OeSNN). The OeSNN offers a robust framework for knowledge discovery in streaming data since it can evolve and adapt to new data patterns in real-time, thereby eliminating the need for retraining. However, reliance on manual hyperparameter tuning presents notable challenges for OeSNN that can compromise model accuracy and stability. To address these challenges, this work introduces a novel hybrid approach (HABCOeSNN) which combines the Artificial Bee Colony (ABC) algorithm with Online Evolving Spiking Neural Networks (OeSNN). The HABCOeSNN extensively investigates the optimization of five key hyperparameters, including window size (Wsize), anomaly classification factor (ε), similarity value (SIM), modulation factor (MOD), and threshold factor (C). The proposed method was thoroughly evaluated on two benchmark datasets, the Numenta Anomaly Benchmark (NAB) and Yahoo Webscope, using a comprehensive set of evaluation metrics. To ensure a more accurate evaluation, we employed a Multi-Criteria Decision-Making (MCDM) approach to validate performance across multiple criteria, rather than relying on a single metric. Further comparative analyses were conducted against five well-established optimization algorithms: Particle Swarm Optimization (PSO), Flower Pollination Algorithm (FPA), Grey Wolf Optimization (GWO), Whale Optimization Algorithm (WOA), and Grid Search (GS), as well as traditional classifiers, Random Forest (RF), Support Vector Machine (SVM), and k-Nearest Neighbors (k-NN). An ablation study was performed to assess the individual contributions of the HABCOeSNN components. The experimental results demonstrate that HABCOeSNN achieves superior performance across diverse time series, with F1-scores ranging from 0.749 to 0.942 on Yahoo Webscope and from 0.427 to 0.878 on the NAB. The results reveal that HABCOeSNN consistently outperforms a wide range of baselines in terms of accuracy and reliability, as confirmed by a statistical one-way analysis of variance (ANOVA) test. These findings highlight the crucial role of automated hyperparameter optimization in improving the performance of OeSNN for unsupervised anomaly detection in streaming data environments.

Nowadays, anomaly detection in streaming data has gained considerable attention due to the exponential growth in the data gathered by Internet of Things applications. Analyzing and processing vast data volumes requires a system capable of working in real-time. Moreover, obtaining labeled data for supervised learning is challenging, as it requires human involvement, is time-consuming, and costly. A promising direction is implementing evolving spiking neural networks (eSNN), which can be updated whenever new data becomes available without re-training previous samples. However, eSNN encounters significant challenges when it comes to manually tuning its hyperparameter values. As such, this work covers the current research gap by suggesting a novel method to optimize the hyperparameters of eSNN called online evolving spiking neural networks-artificial bee colony (OeSNN-ABC). Multiple scenarios have been utilized to evaluate the proposed method using two benchmark datasets: the Numenta anomaly benchmark (NAB) and the Yahoo Webscope using different criteria. Further validation was provided by comparing the proposed OeSNN-ABC against five well-known optimization algorithms: particle swarm optimization, grey wolf optimization, flower pollination algorithm, whale optimization algorithm, and grid search, alongside other classifiers such as random forest, support vector machine, and k-nearest neighbor. The findings revealed that OeSNN-ABC had the best performance among all compared optimization algorithms and classifiers, outperformed prior anomaly detection techniques for the NAB dataset, and achieved competitive results for the Yahoo Webscope dataset.