The generation of content by Artificial Intelligence (AI) in educational settings represents a growing consequence of the irresponsible use of generative AI that threatens academic integrity. Mitigating this phenomenon through AI-based detection remains a major challenge, given that generative AI tools, such as Large Language Models (LLMs), evolve rapidly. Classical detection tools such as anti-plagiarism software are no longer capable of tracking or curbing the spread of this phenomenon. Furthermore, the existing AI-based tools fail to detect the content generated, paraphrased, or humanized by bots that alter sentence structure. In this article, we propose a novel architecture based on Generative Adversarial Networks (GANs). This framework is a generative model composed of two main components trained through a minimax game. The first is the generator; it generates or reformulates text, while its opponent, the discriminator, tries to detect whether the text is human-generated or produced by the generator. We used DistilGPT-2 as the generator, trained simultaneously with DistilRoBERTa, the discriminator. Our framework was evaluated on 44,868 academic text samples and achieved 99% precision on unmodified texts. Moreover, our approach demonstrates strong reliability, maintaining a precision of 92% despite paraphrasing and humanization, as well as noise injection attacks that render other detection tools ineffective. The use of GANs enables our discriminator to more effectively identify deep semantic patterns and reformulated structures. Finally, we discuss future research directions to improve the adaptability and long-term relevance of the proposed approach in light of rapid advances in artificial intelligence.
Algorithms (GA), developed several decades ago, are widely used to solve optimization problems. Since the invention of genetic algorithms, significant progress has been made in genetics, leading to the discovery of many new mechanisms of heredity that were not previously incorporated into genetic algorithms. One such group is epigenetic processes. The article presents two epigenetic-based operations used in genetic algorithms: cytosine methylation and allele inactivation (gene silencing). This article presents two original versions of genetic algorithms that use epigenetic processes. It presents the results of research on the impact of the introduced modifications on the genetic algorithm's performance. The study examined two optimization problems: the loading optimization problem and the outlier detection problem. The experiments examined the optimal probability of occurrence of the introduced modifications, the average number of iterations of the genetic algorithm with and without modifications required to obtain the optimal result, and compared execution times. The research confirmed that introducing new operations into genetic algorithms can improve their performance.
ripeness of oil palm fresh fruit bunches significantly impacts the quality and economic value of palm oil production. This study proposes a novel approach using to classify and detect the ripeness of oil palm fruit bunches. By leveraging pretrained models and fine-tuning techniques, this study aims to improve detection accuracy while reducing inference time. The methodology involves data preparation, model training, and evaluation, utilizing a dataset comprising 427 images of oil palm fruit bunches classified into raw, halfripe, and ripe categories. Additionally, the images were augmented to increase the dataset size. The YOLOv8 architecture, known for its scalability and efficiency, is applied to improve the classification process. Results demonstrate that YOLOv8 provides a balance between accuracy and processing speed, making it a suitable tool for real-time applications in the palm oil industry. This study contributes to reducing the reliance on manual techniques, lowering operational costs, and increasing the overall efficiency of the oil palm industry.
Operating Room (OR) efficiency directly influences patient safety, resource allocation, and overall hospital performance. Despite the routine monitoring of operating room quality indicators, most analyses remain retrospective and provide limited support for forward-looking management. This study develops a forecasting framework based on Long Short-Term Memory (LSTM) neural networks and integrates the prediction results into a Business Intelligence (BI) visualization environment. Monthly specialty-level operational data from a medical center in Taiwan covering 2021 to 2024 were analyzed, including seven key operating room quality indicators. Different historical input windows ranging from 3 to 24 months were examined to evaluate their effects on predictive performance. To assess whether deep learning was necessary for this task, model performance was compared with traditional statistical approaches, including Moving Average, automated Autoregressive Integrated Moving Average (auto-ARIMA), and Na & iuml;ve forecasting models, using a strictly chronological training and testing split. Indicators characterized by short-term operational variability, such as cancellation and scheduling delay rates, were more accurately predicted using shorter historical windows, whereas structurally stable metrics including utilization and occupancy rates showed improved performance with longer input sequences. While conventional statistical models performed competitively for some stable indicators, LSTM provided more consistent performance across heterogeneous metrics. The forecasting outputs were incorporated into an interactive dashboard to facilitate real-time monitoring and threshold-based management. Although the analysis was conducted using data from a single medical center, the results demonstrate the feasibility of integrating predictive analytics into routine surgical quality management.
Data collection in animal ecology has become more common, and extracting information can be done using deep learning methods. Challenges arise when using the dataset as the training input, including balancing the data quality and quantity. This research aims to investigate the effect of data balancing techniques on the optimizer-specifically, Root Mean Square Propagation (RMSProp) and Adam-in a deep learning image recognition model. Once recognition capability is obtained, it provides feedback for an Image-Based Visual Servoing (IBVS) system. The wildlife dataset is from Serengeti National Park and contains 41,762 images across 7 classes. Before applying data-balancing techniques, we determined that 1,100 images were needed, which we considered a balanced distribution across all classes since it was close to the original dataset. The next step was applying the balancing technique. We offered two balancing techniques: the first was to augment three classes with each class's 1,100 images, while the other four classes retained their original number of images. The second involved reducing the number of larger classes to 1,100 images. The result showed that combining the deep learning model AlexNet with RMSProp achieved the highest Macro F1-Score of 0.8118, outperforming the other deep learning models with the Adam optimizer. The best balancing technique reduced the larger class to a fixed number of images in the smaller dataset, and AlexNet with RMSProp is a good choice for realtime visual servoing.
diseases remain the leading cause of mortality worldwide, emphasizing the need for accurate early prediction. With the increased accessibility to clinical datasets, Machine Learning (ML)-classification algorithms and Deep Learning (DL) approaches (Convolutional Neural Networks) have become integral to early prediction of cardiovascular diseases. This work proposes an effectual hyperparameter optimized heart disease prediction system using a Multilayered Perceptron (MLP) and compares its performance with other Machine Learning (ML) algorithms. The Cleveland, Statlog and Hungarian heart disease datasets sourced from the UC Irvine Machine Learning Repository (UCI ML) were used. Initially, data pre-processing includes correlation analysis between the relevant body values and disease-followed by applying Synthetic Minority Over sampling Technique (SMOTE) on the training folds only to address the class imbalances without affecting the data integrity. A 10-Fold Cross-Validation approach was used to ensure the robust model evaluation. Hyperparameter optimization using GridsearchCV further enhanced the generalization. Optimized MLP attained the highest accuracy levels of 96.89 followed by the Random Forest (96.07%), Decision Tree (94.99%), K-Nearest Neighbor (KNN) (94.10%), Support Vector Machines (SVM) (93.48%) and Logistic Regression (LR) (93.55%). The Findings obtained illustrate the efficacy of MLP architecture in the prediction of cardio vascular diseases for clinical applications.
Structural cracks are commonly caused by factors such as high loading, fatigue, degradation, thermal expansion, and humidity. Detection of cracks in early stages is important. If neglected, these cracks can spread and weaken the building as well as other civil infrastructure. To enhance safety and prevent catastrophic structural failures, accurately detecting cracks is essential in Structural Health Monitoring (SHM) systems, thus enabling early maintenance. The latest developments in Deep Learning (DL) have facilitated the automatic detection of structural cracks through images taken with the help of cameras, drones, or mobile devices. Such methods minimize the manual inspection and enhance monitoring efficiency. In this paper, a systematic optimization of deep learning models through hyperparameter tuning with random search was implemented to achieve a balance between classification performance and computational efficiency for resource-constrained environments. Custom architecture, (Convolutional Neural Network (CNN) and CNN-Long Short-Term Memory (LSTM)), as well as transfer learning models (MobileNetV2, ResNet101, and DenseNet201), are evaluated on the Middle East Technical University (METU) concrete crack dataset (40,000 labelled images). The performance measures, such as accuracy, precision, recall, F1-Score, inference time, and computational complexity in Million Floating-Point Operations (MFLOPs), are evaluated and compared on the models. From the results, MobileNetV2 shows a classification accuracy of 99.8% with a reduction of 50% computational complexity, thus making it suitable for deployment on constrained devices in real-world applications.
Time series forecasting plays a pivotal role in domains ranging from finance and healthcare to multimodal analytics, yet achieving robust accuracy across heterogeneous datasets and varying prediction horizons remains a significant challenge. In this paper, we propose a novel hybrid forecasting framework that integrates advanced metaheuristic optimization with deep learning architectures for multi-domain, multi-scale prediction. The core of our approach is a dual-stage optimization strategy that combines the global exploration capability of the Aquila Optimizer (AO) with the guided local refinement of the Proposed Guided Remora Optimization Algorithm (PGRO). This hybrid optimization mechanism is applied to fine-tune model hyperparameters, fusion strategies, and incremental learning settings for architectures such as Long Short Term Memory (LSTM), Transformer, and multimodal fusion networks. The framework is further enhanced by an incremental learning module with memory replay, enabling continual adaptation without catastrophic forgetting. We evaluate the proposed approach on 6 diverse datasets (National Association of Securities Dealers Automated Quotations/New York Stock Exchange (NASDAQ/NYSE) Historical Stock, Financial Question Answering (FiQA), Financial PhraseBank, Reddit Financial News, Fourth Makridakis Forecasting Competition Time series (M4), and Video and Text (VaTeX) Multimodel), spanning structured numerical, textual, social media, and multimodal time series. Experimental results show that our AO + PGRO-enhanced models consistently outperform classical statistical models, deep learning baselines, and state-of-the-art transformer-based forecasters in both short-and long-horizon prediction tasks. Ablation studies confirm the contribution of each component, and detailed horizon-wise analysis highlights the method's adaptability to varying temporal dependencies. The proposed framework delivers state-of-the-art accuracy, stability, and generalization across domains, offering a powerful and extensible approach for real-world time series forecasting challenges.
Automated Essay Scoring (AES) systems have achieved notable success in evaluating semantic content using deep learning models. However, they often fail to provide explicit, interpretable feedback on essay structure. This paper presents a hybrid neural-symbolic approach that integrates Bidirectional Encoder Representations from Transformers (BERT) for semantic analysis with a Hidden Markov Model (HMM) for explicit structural modeling. Unlike ensemble methods, our system employs a tightly coupled, weighted integration scheme (70% content, 30% structure) optimized through validation experiments. The HMM component models essay organization as a sequence of rhetorical states Introduction, Body, and Conclusion offering transparent feedback on logical flow. We explicitly position this work as a methodological proof-of-concept validation study.
This work presents a portable, modular quantum-circuit emulator designed to accelerate algorithm prototyping by leveraging Central Processing Unit (CPU) and Compute Unified Device Architecture (CUDA)-enabled Graphic Processing Unit (GPU) backends for dense linear-algebra workloads. The emulator provides a lightweight backend abstraction that unifies NumPy and CuPy, supports both statevector and density-matrix representations, and implements efficient multi-target unitary embedding and Kraus-map noise channels. These capabilities enable rapid exploration of quantum algorithms that require full 2n & times; 2n operators, such as quantum phase estimation, full Quantum Fourier Transform (QFT), and noise-aware variational circuits, while minimizing host/device transfer and preserving numerical parity between backends. Implementation strategies are described for embedding single-and multi-qubit gates, parameterized rotations, and controlled unitaries, along with an "apply_unitary_on_targets" routine that employs reshape/matrix-multiply/reshape patterns to map arbitrary target sets to dense GPU kernels for high throughput. Using a curated benchmark suite of dense quantum workloads and microbenchmarks, including large matrix multiplies and Fast Fourier Transforms (FFT), performance is quantified across a range of qubit counts and precisions. The results demonstrate substantial speedups on modern NVIDIA GPUs for workloads dominated by dense linear algebra, identify cross-over points where GPU acceleration becomes beneficial, and analyse memory/precision trade-offs for complex64 versus complex128. The presented emulator lowers the barrier for noise-aware algorithm prototyping by combining practical software ergonomics with GPU performance, enabling faster iteration on algorithm design and noise-mitigation strategies on commodity single-node hardware. While the abstraction layer ensures portability, achieving peak performance currently requires hardware-specific tuning-a limitation we address through future automation strategies.
