18 sonuçlar
Arama Sonuçları
Listeleniyor 1 - 10 / 18
Yayın Electroencephalography signatures associated with developmental dyslexia identified using principal component analysis(Multidisciplinary Digital Publishing Institute (MDPI), 2025-08-27) Eroğlu, Günet; Harb, Mhd Raja AbouBackground/Objectives: Developmental dyslexia is characterised by neuropsychological processing deficits and marked hemispheric functional asymmetries. To uncover latent neurophysiological features linked to reading impairment, we applied dimensionality reduction and clustering techniques to high-density electroencephalographic (EEG) recordings. We further examined the functional relevance of these features to reading performance under standardised test conditions. Methods: EEG data were collected from 200 children (100 with dyslexia and 100 age- and IQ-matched typically developing controls). Principal Component Analysis (PCA) was applied to high-dimensional EEG spectral power datasets to extract latent neurophysiological components. Twelve principal components, collectively accounting for 84.2% of the variance, were retained. K-means clustering was performed on the PCA-derived components to classify participants. Group differences in spectral power were evaluated, and correlations between principal component scores and reading fluency, measured by the TILLS Reading Fluency Subtest, were computed. Results: K-means clustering trained on PCA-derived features achieved a classification accuracy of 89.5% (silhouette coefficient = 0.67). Dyslexic participants exhibited significantly higher right parietal–occipital alpha (P8) power compared to controls (mean = 3.77 ± 0.61 vs. 2.74 ± 0.56; p < 0.001). Within the dyslexic group, PC1 scores were strongly negatively correlated with reading fluency (r = −0.61, p < 0.001), underscoring the functional relevance of EEG-derived components to behavioural reading performance. Conclusions: PCA-derived EEG patterns can distinguish between dyslexic and typically developing children with high accuracy, revealing spectral power differences consistent with atypical hemispheric specialisation. These results suggest that EEG-derived neurophysiological features hold promise for early dyslexia screening. However, before EEG can be firmly established as a reliable molecular biomarker, further multimodal research integrating EEG with immunological, neurochemical, and genetic measures is warranted.Yayın Extracting meaningful information student surveys with NLP(Işık Üniversitesi, Lisansüstü Eğitim Enstitüsü, 2025-01-29) Pourjalil, Kajal; Ekin, Emine; Işık Üniversitesi, Lisansüstü Eğitim Enstitüsü, Bilgisayar Mühendisliği Yüksek Lisans Programı; Işık University, School of Graduate Studies, Master’s Program in Computer EngineeringThis thesis applied NLP techniques to analyze and summarize bilingual student feedback collected via end-of-semester surveys. The dataset, which contained open-ended responses in both English and Turkish, required a model adept at preserving linguistic nuances across languages. The Llama 2-7b-hf model, which had been trained explicitly for text generation, was selected for its capability to produce coherent and contextually relevant summaries. Data preprocessing involved organizing metadata such as department, semester, course name, and section number, segregating comments by word count, and removing personal identifiers to ensure privacy. Shorter comments (fewer than ten words) were grouped and summarized using a pipeline from the Transformers library, while longer comments were fine-tuned with metadataspecific prompts for detailed summarization. To further enhance analysis, sentiment classification was performed using the “cardiffnlp/twitter-robertabase-sentiment” model, categorizing feedback into negative, neutral, and positive sentiments. Evaluation metrics included expert reviews, contextual relevance, and logical consistency with the dataset’s sentiment distribution. Compared to previous models, the Llama 2 model demonstrated superior performance in generating complete, coherent summaries while preserving the overall intent and tone of the comments. Ultimately, this research highlighted the effectiveness of LLMs in processing multilingual educational data and their potential to provide actionable insights for improving course content and student experiences.Yayın Image super resolution using deep learning techniques(Işık Üniversitesi, Lisansüstü Eğitim Enstitüsü, 2024-09-02) El Ballouti, Salah Eddine; Eskil, Mustafa Taner; Işık Üniversitesi, Lisansüstü Eğitim Enstitüsü, Bilgisayar Mühendisliği Yüksek Lisans Programı; Işık University, School of Graduate Studies, Master’s Program in Computer EngineeringImage SR using Deep Learning Techniques has become a critical area of research, with significant progress in improving image quality and detail. This thesis examines and contrasts eight advanced deep learning-based SR methods: CARN, EDSR, ESPCN, RCAN, RDN, SRCNN, SRGAN, and VDSR, using the DIV2K dataset. The evaluation covers multiple aspects to offer a thorough understanding of each method's effectiveness, efficiency, and structure. Performance measurements such as PSNR and SSIM are utilized for evaluating the fidelity of super-resolved images. Computational efficiency is evaluated based on inference time and memory requirements. Training time is analyzed, taking into account the speed of convergence for training on the DIV2K dataset. Model complexity is examined, exploring architectural details such as network depth, and the integration of specialized elements like residual blocks and attention mechanisms. Additionally, the thesis explains in a clear and detailed manner the trade-offs between performance and complexity, discussing whether more complex architectures deliver significantly better results compared to simpler models and whether the computational cost justifies the improvements. Finally, a qualitative comparison is conducted to emphasize the strengths and weaknesses of each technique. Through this comprehensive analysis, this thesis offers insights into the field of deep learning-based image SR, assisting researchers and practitioners in choosing the most appropriate method for various applications.Yayın Efficient estimation of Sigmoid and Tanh activation functions for homomorphically encrypted data using Artificial Neural Networks(Institute of Electrical and Electronics Engineers Inc., 2024) Harb, Mhd Raja Abou; Çeliktaş, BarışThis paper presents a novel approach to estimating Sigmoid and Tanh activation functions using Artificial Neural Networks (ANN) optimized for homomorphic encryption. The proposed method is compared against second-degree polynomial and Piecewise Linear approximations, demonstrating a minor loss in accuracy while maintaining computational efficiency. Our results suggest that the ANN-based estimator is a viable alternative for secure machine learning models requiring privacypreserving computation.Yayın A context-aware, AI-driven load balancing framework for incident escalation in SOCs(Institute of Electrical and Electronics Engineers Inc., 2025-08-12) Abuaziz, Ahmed; Çeliktaş, BarışSOCs face growing challenges in incident management due to increasing alert volumes and the complexity of cyberattacks. Traditional rule-based escalation models often fail to account for the workload of the analyst, the severity of the incident, and the organizational context. This paper proposes a context-aware, AI-driven load balancing framework for intelligent analyst assignment and incident escalation. Our framework leverages large language models (LLMs) with retrievalaugmented generation (RAG) to evaluate incident relevance and historical assignments. A reinforcement learning (RL)-based scheduler continuously optimizes incident-to-analyst assignments based on operational outcomes, enabling the system to adapt to evolving threat landscapes and organizational structures. Planned simulations in realistic SOC environments will compare the model with traditional rule-based models using metrics such as Mean Time to Resolution (MTTR), workload distribution, and escalation accuracy. This work highlights the potential of AIdriven approaches to improve SOC performance and enhance incident response effectiveness.Yayın Federated hybrid privacy-preserving movie recommendation system for internet-of-vehicles(Işık Üniversitesi, Lisansüstü Eğitim Enstitüsü, 2024-08-02) Şimşek, Musa; Erman Tüysüz, Ayşegül; Işık Üniversitesi, Lisansüstü Eğitim Enstitüsü, Bilgisayar Mühendisliği Yüksek Lisans Programı; Işık University, School of Graduate Studies, Master’s Program in Computer EngineeringIn this research, we introduced a pioneering strategy to address the pressing privacy concerns associated with vehicular movie recommendation systems. As the demand for personalized entertainment options in vehicles increases, so does the need to protect user data. To tackle this challenge, we utilized the PyTorch framework to create a robust foundation from scratch. A key component of our approach was the addition of Laplace noise during the training process, which ensured differential privacy. This technique effectively safeguarded user data while simultaneously optimizing model performance, allowing us to maintain high levels of recommendation accuracy. Furthermore, we employed the Optuna hyperparameter optimization framework, which played a crucial role in enhancing the model's performance. By fine-tuning various parameters, we were able to elevate the overall efficiency of the system beyond the capabilities of the base model. Our extensive experimentation utilized the Movielens-1M benchmark movie dataset, which provided a solid basis for evaluating our approach. The results demonstrated a significant improvement over baseline models, validating the effectiveness of our privacy-preserving vehicular movie recommendation system. In addition to our centralised model, we conducted a comprehensive comparison with practical federated frameworks, including FedAvg, FedProx, and FedMedian. Our findings revealed that all federated models outperformed the centralised models by at least 2%, while also exhibiting shorter runtimes.Yayın ANN activation function estimators for homomorphic encrypted inference(Institute of Electrical and Electronics Engineers Inc., 2025-06-13) Harb, Mhd Raja Abou; Çeliktaş, BarışHomomorphic Encryption (HE) enables secure computations on encrypted data, facilitating machine learning inference in sensitive environments such as healthcare and finance. However, efficiently handling non-linear activation functions, specifically Sigmoid and Tanh, remains a significant computational challenge for encrypted inference using Artificial Neural Networks (ANNs). This study introduces a lightweight, ANN-based estimator designed to accurately approximate activation functions under homomorphic encryption. Unlike traditional polynomial and piecewise linear approximations, the proposed ANN estimators achieve superior accuracy with lower computational overhead associated with bootstrapping or high-degree polynomial techniques. These estimators are trained on plaintext data and seamlessly integrated into encrypted inference pipelines, significantly outperforming conventional methods. Experimental evaluations demonstrate notable improvements, with ANN estimators enhancing accuracy by approximately 2% for Sigmoid and up to 73% for Tanh functions, improving F1-scores by approximately 2% for Sigmoid and up to 88% for Tanh, and markedly reducing Mean Square Error (MSE) by up to 96% compared to polynomial approximations. The ANN estimator achieves an accuracy of 97.70% and an AUC of 0.9997 when integrated into a CNN architecture on the MNIST dataset, and an accuracy of 85.25% with an AUC of 0.9459 on the UCI Heart Disease dataset during ciphertext inference. These results underscore the estimator’s practical effectiveness and computational feasibility, making it suitable for secure and efficient ANN inference in encrypted environments.Yayın Intelligent health monitoring in 6G networks: machine learning-enhanced VLC-based medical body sensor networks(Multidisciplinary Digital Publishing Institute (MDPI), 2025-05-23) Antaki, Bilal; Dalloul, Ahmed Hany; Miramirkhani, FarshadRecent advances in Artificial Intelligence (AI)-driven wireless communication are driving the adoption of Sixth Generation (6G) technologies in crucial environments such as hospitals. Visible Light Communication (VLC) leverages existing lighting infrastructure to deliver high data rates while mitigating electromagnetic interference (EMI); however, patient movement induces fluctuating signal strength and dynamic channel conditions. In this paper, we present a novel integration of site-specific ray tracing and machine learning (ML) for VLC-enabled Medical Body Sensor Networks (MBSNs) channel modeling in distinct hospital settings. First, we introduce a Q-learning-based adaptive modulation scheme that meets target symbol error rates (SERs) in real time without prior environmental information. Second, we develop a Long Short-Term Memory (LSTM)-based estimator for path loss and Root Mean Square (RMS) delay spread under dynamic hospital conditions. To our knowledge, this is the first study combining ray-traced channel impulse response modeling (CIR) with ML techniques in hospital scenarios. The simulation results demonstrate that the Q-learning method consistently achieves SERs with a spectral efficiency (SE) lower than optimal near the threshold. Furthermore, LSTM estimation shows that D1 has the highest Root Mean Square Error (RMSE) for path loss (1.6797 dB) and RMS delay spread (1.0567 ns) in the Intensive Care Unit (ICU) ward, whereas D3 exhibits the highest RMSE for path loss (1.0652 dB) and RMS delay spread (0.7657 ns) in the Family-Type Patient Rooms (FTPRs) scenario, demonstrating high estimation accuracy under realistic conditions.Yayın Assessing ChatGPT's accuracy in dyslexia inquiry(Institute of Electrical and Electronics Engineers Inc., 2024) Eroğlu, Günet; Harb, Mhd Raja AbouDyslexia poses challenges in accessing reliable information, crucial for affected individuals and their families. Leveraging chatbot technology offers promise in this regard. This study evaluates the OpenAI Assistant's precision in addressing dyslexia-related inquiries. Three hundred questions commonly posed by parents were categorized and presented to the Assistant. Expert evaluation of responses, graded on accuracy and completeness, yielded consistently high scores (median=5). Descriptive questions scored higher (average=4.9568) than yes/no questions (average=4.8957), indicating potential response challenges. Statistical analysis highlighted the significance of question specificity in response quality. Despite occasional difficulties, the Assistant demonstrated adaptability and reliability in providing accurate dyslexia-related information.Yayın “Can we use a biomarker detection algorithm to measure the effectiveness of 14-channel neurofeedback in dyslexia?”(Routledge, 2025-10-01) Eroğlu, Günet; Harb, Raja AbouDyslexia, one of children’s most common neurological diversities, primarily manifests as a reduced reading ability. Genetic factors contribute to dyslexia, with contemporary theories attributing it to a delay in left hemispheric lateralization that reduces effective reading and writing skills. To assist dyslexic children, smartphone application, Auto Train Brain, has been developed to enhance reading comprehension and speed. Previously, the efficacy of the mobile application’s training program was assessed using psychometric tests; however, our study employed a biomarker detection software to evaluate the neurofeedback’s impact. Machine learning (ML) techniques have recently gained traction in differentiating between dyslexia and typically developing children (TDC). The dataset of this study consists of 100 sessions of 2-minute resting-state eyes-open 14-channel Quantitative Electroencephalography (QEEG) data from 100 children with dyslexia and 100 TDC. Therefore, the dyslexia biomarker detection software assessed the efficacy of the 14-channel neurofeedback administered via Auto Train Brain. Results showed significant improvement in electrophysiological normalization, increasing from 30% in the first 20 sessions to 61% by the end of the training. A two-proportion Z-test confirmed this improvement was statistically significant (Z = −3.96, p = 0.00007), particularly between the 1–20 and 1–60 session intervals (Z = −2.66, p = 0.0079).
