3 sonuçlar
Arama Sonuçları
Listeleniyor 1 - 3 / 3
Yayın Uyarlanır yerel bağlı nöron modelinin incelemesi(Gazi Üniversitesi Bilişim Enstitüsü, 2019-04-10) Tek, Faik BorayBu çalışmada uyarlanır yerel bağlı (odaklanan) nöron modelinin bir incelemesi sunulmuştur. Öncelikle bu modelin varolan diğer nöron modelleri ile ilişkisi incelenmiştir. Daha sonra modelin ileri beslemede çalışması ve geriye yayılım ile eğitilmesi tartışılmıştır. Modelin çalışma prensipleri sentetik sınıflandırma veri kümeleri üzerinde deneylerle gösterilmiştir. Son olarak, basit ve evrişimli ağların saklı katmanlarında odaklı nöronlar kullanılması halinde tam bağlı nöronlara göre daha iyi bir performans elde edilebileceği MNIST, CIFAR10, FASHION gibi popüler imge tanıma veri kümelerinde karşılaştırmalı olarak gösterilmiştir.Yayın Electrical circuit design based on neural networks(Işık Üniversitesi, Lisansüstü Eğitim Enstitüsü, 2026-01-23) Abou Allil, Feras; Köprü, Ramazan; Işık Üniversitesi, Lisansüstü Eğitim Enstitüsü, Elektrik-Elektronik Mühendisliği Yüksek Lisans Programı; Işık University, School of Graduate Studies, Electric-Electronics Engineering M.S. ProgramArtificial Neural Networks (ANNs) have gained significant attention due to their fast and accurate performance estimation capabilities, particularly in applications requiring strong learning and generalization. In this thesis, a comprehensive study is presented on the use of neural networks for the design and analysis of analog electronic circuits, focusing on both passive and active filter topologies. A feedforward neural network architecture is employed to reduce unwanted noise in measurement signals and to accurately infer component values from frequency response characteristics. For each circuit type, a dedicated neural network is trained to learn the relationship between circuit parameters and their corresponding magnitude responses. The study includes a variety of analog filters—such as low-pass and band-pass filters—implemented using passive elements as well as active devices including operational amplifiers and operational transconductance amplifiers (OTAs). Two training methodologies are introduced and evaluated: Element Spreading Training (EST) and Element Randomization Training (ERT). These approaches enhance dataset diversity and improve the neural network’s ability to generalize across a wider range of circuit behaviors, resulting in more reliable and robust predictions. The overall framework demonstrates the potential of integrating neural networks into classical analog circuit design, offering insights into performance, advantages, and limitations. All analyses and simulations are conducted and validated using MATLAB. The proposed methods have been tested under different frequency ranges and component tolerances.Yayın Advancing privacy and security in machine learning through homomorphic encryption and explainable AI(Işık Üniversitesi, Lisansüstü Eğitim Enstitüsü, 2026-03-05) Abou Harb, Mhd Raja; Çeliktaş, Barış; Işık Üniversitesi, Lisansüstü Eğitim Enstitüsü, Bilgisayar Mühendisliği Doktora Programı; Işık University, School of Graduate Studies, Ph.D. in Computer EngineeringThe importance of data privacy in cloud-based Machine Learning is paramount, particularly in sectors such as healthcare and finance. Balancing robust privacy protection with high model accuracy remains a significant challenge. In this study, we propose a privacy-preserving framework utilizing ANNs on homomorphically encrypted data. To mitigate the computational complexity of non-linear activation functions (Sigmoid and Tanh), we developed lightweight, ANN-based estimators specifically designed for encrypted environments. Our experimental results demonstrate that these estimators significantly outperform traditional polynomial and piecewise linear methods, reducing MSE by up to 96% while improving accuracy and F1-scores. Our method achieved 97.70% accuracy and 0.9997 AUC on the MNIST dataset, validating its effectiveness. In real-world applications, we applied the approach to dyslexia detection using QEEG data, observing only minor performance degradation (2.66% accuracy, 3.86% AUC) compared to plaintext inference. Furthermore, a case study on the UCI Heart Disease dataset yielded 85.25% accuracy in encrypted inference, matching plaintext performance. Finally, we integrated the SHAP algorithm to ensure transparency for encrypted outputs. Our findings confirm that this approach successfully balances privacy, performance, and explainability, making it highly suitable for sensitive ML applications.
