LEE- Hesaplamalı Bilim ve Mühendislik-Yüksek Lisans
Bu koleksiyon için kalıcı URI
Gözat
Son Başvurular
1 - 5 / 12
-
ÖgeComparison of truncation and otsu-based thresholds for decoding subthreshold motor cortical activity in extracellular neural recordings(Graduate School, 2025-06-30)This study examines two automated methods that enable the separation of extracellular neural recordings filtered in the action potential band into signal and subthreshold components known as ’noise’: Truncation Thresholds and Otsu-based Thresholds. While traditional methods generally focus on suprathreshold data, Truncation Thresholds and Otsu-based Thresholds methods offer a novel approach to behavioral decoding by utilizing subthreshold data. Previous studies have shown that the features (mean (µ) and standard deviation (σ)) of subthreshold data identified by Truncation Thresholds change with a rat’s behavior of pressing either the right or left pedal, and behavior can be decoded with up to 100% accuracy using these features. It has also been determined that Otsu-based Thresholds estimate the standard deviation of subthreshold components in simulated data more accurately than Truncation Thresholds. It implies that Otsu-based Thresholds method provides a more robust estimation by remaining unaffected by increasing neuronal firing rates.
-
ÖgeComression of convolutional neural networks via high dimensional model representation(ITU Graduate School, 2025)In recent years, the growing complexity of deep learning architectures—particularly convolutional neural networks (CNNs)—has introduced substantial challenges related to model size, computational overhead, and real-time deployability on resource-limited platforms. Although architectures such as ResNet and VGG have set new benchmarks in image classification tasks by leveraging deep and expressive structures, their significant number of parameters and high inference latency present obstacles when deploying these models on edge devices, embedded systems, or mobile applications. These challenges underscore the necessity of effective model compression techniques that maintain accuracy while reducing computational and memory demands. This thesis addresses these limitations by systematically exploring tensor-based model compression strategies for CNNs, with a specific focus on Tensor Train (TT) decomposition and High Dimensional Model Representation (HDMR). TT decomposition is a powerful low-rank tensor factorization technique that compresses high-dimensional weight tensors into a sequence of interconnected low-rank cores, significantly reducing the parameter count while preserving the structural hierarchy of convolutional layers. HDMR, in contrast, is a functional decomposition approach that approximates a multivariate function as a hierarchical sum of univariate and multivariate component functions, each capturing individual or interactive effects of input variables. While TT operates on the structural representation of tensors, HDMR operates in the functional domain, enabling interpretability and capturing non-linear interactions. Beyond analyzing these techniques in isolation, the study also introduces and evaluates two hybrid compression strategies: (1) TT->HDMR, where TT decomposition is first applied to CNN weight tensors, followed by HDMR analysis on the resulting TT cores, and (2) HDMR->TT, where HDMR is initially used to functionally decompose the weight tensors, and TT is subsequently applied to compress the resulting components. These methods are integrated into four popular CNN architectures—ResNet-18, ResNet-34, VGG16, and VGG19—and evaluated on the CIFAR-10 dataset. Experimental results reveal that TT->HDMR consistently achieves favorable compression-accuracy trade-offs. This strategy leads to up to 6.31× parameter reduction in certain configurations (e.g., ResNet-34) while preserving or even improving the classification accuracy. In contrast, HDMR->TT often results in increased parameter counts and degraded performance due to the high number of interaction terms generated during HDMR expansion (especially when higher-order terms such as 3rd-order components are included). When HDMR is applied directly, a moderate compression is observed, but this comes at the cost of increased inference time and memory due to the expansion of functional components. These findings highlight the importance of decomposition order and the selection of HDMR truncation order in determining the efficiency and effectiveness of hybrid compression strategies. Furthermore, the thesis shows that TT alone is highly effective in compressing convolutional layers while maintaining a balanced performance profile. It is particularly well-suited for models with redundant parameter structures such as VGG19. HDMR, while not inherently a tensor-based method, introduces new perspectives in interpretable compression, although its direct integration into deep networks must be done cautiously to avoid parameter inflation. In conclusion, this study does not merely promote one optimal solution but instead explores a space of tensor-based and function-based compression methods with the aim of discovering whether more efficient decompositions can be achieved without sacrificing accuracy. The results demonstrate that hybrid methods—when applied with the right sequence and order truncation—can uncover more compact and accurate representations of deep CNNs, suitable for real-world deployment. These findings offer valuable guidance for future research, including potential adaptations of TT and HDMR techniques to transformer-based architectures, federated learning settings, or multi-modal models where compression, communication cost, and interpretability are all critical factors.
-
ÖgeDoğal dil ile SQL ve görselleştirme koduna dönüşümde büyük dil modellerinin karşılaştırmalı analizi(İTÜ Lisansüstü Eğitim Enstitüsü, 2025)Veriye erişimin kolaylaştırılması ve teknik olmayan kullanıcılar için veri tabanlarıyla etkileşimin demokratikleştirilmesi, günümüz bilgi teknolojileri açısından büyük önem arz etmektedir. Bu bağlamda, doğal dil sorgularını yapılandırılmış SQL ifadelerine ve görsel Python kodlarına dönüştürebilen büyük dil modelleri (LLM – Large Language Models), geleneksel veri analizi süreçlerine önemli katkılar sunmaktadır. Bu tez çalışmasında, doğal dil ile veri sorgulama ve görselleştirme süreçlerini uçtan uca gerçekleştiren bir yapay zekâ destekli sistem tasarlanmış ve farklı büyük dil modellerinin bu sistem üzerindeki performansları karşılaştırmalı olarak analiz edilmiştir. Çalışma kapsamında OpenAI GPT-4o, Anthropic Claude 3.5 Sonnet, Meta LLaMA 3.1 70B, Google Gemini 1.5 Flash ve DeepSeek Chat modelleri değerlendirilmiştir. Microsoft'un açık kaynaklı Semantic Kernel yazılım geliştirme kiti kullanılarak doğal dil sorgularının SQL ve Python kodlarına dönüştürüldüğü modüler ve genişletilebilir bir sistem mimarisi geliştirilmiştir. Sistem, kullanıcıdan gelen doğal dil girdisini şema açıklaması ile birlikte işleyerek çalıştırılabilir SQL ve Python kodları üretmekte, ardından bu kodları çalıştırarak sonuçları kullanıcıya sunmaktadır. Model başarımı hem insan değerlendirmesiyle hem de en başarılı modelin hakem olarak kullanıldığı LLM-tabanlı otomatik değerlendirme ile ölçülmüştür. İnsan değerlendirmesine göre Claude 3.5 Sonnet en yüksek doğrulukla çalışırken, GPT-4o ve Gemini Flash da özellikle SQL üretiminde başarılı sonuçlar vermiştir. LLM değerlendirmesinde ise Python görselleştirme kodlarının kalite farkları daha belirgin hale gelmiştir. LLaMA ve DeepSeek modelleri SQL çıktılarında rekabetçi sonuçlar sunarken, Python kod üretiminde daha düşük skorlar almıştır. Bu tez çalışması, farklı büyük dil modellerinin metinden koda dönüşüm yeteneklerini kapsamlı biçimde karşılaştırarak, model seçiminde ve sistem mimarisi kurulumunda yol gösterici olmayı amaçlamaktadır. Ayrıca Semantic Kernel tabanlı yaklaşım, yeni modellerin hızlı entegrasyonuna olanak sağlayan esnek bir altyapı sunmakta ve bu yönüyle sürekli gelişen LLM ekosistemine uyumlu bir çözüm önermektedir.
-
ÖgeDesign of short peptides targeting the interaction between SARS-CoV-2 and human ACE2(Graduate School, 2023)In the end of 2019, a novel coronavirus called SARS-CoV-2 has appeared in Wuhan, China and caused a pandemic outbreak by overwhelming deadly infections around the work. By comparing to its ancestor, SARS-CoV, it is more infectious and shows much more tight binding with host cells. One of the main units of this coronavirus is spike (S) protein which has two subunits called S1 and S2. Subunit S1 contains receptor binding domain (RBD) that takes role in initiating entry into the cells by interacting with human angiotensin converting enzyme 2 (ACE2) receptor. And S2 subunit ensures the fusion between host and viral cell membranes in meantime. Different drug development researches were studied and several therapeutics developed during the pandemic time such as antibodies and vaccines. Differing from traditional drug developments methods, peptides as inhibitor are promising drug compound due to their efficiency, lesser immunogenicity, easiness of removal from body and they have higher diffusivity through tissues and organs because of their smallness. Considering these advantages of peptides, several studies have been made with knowing antiviral peptides by different research groups. In present study, what we aimed to develop novel peptides to use against spike RBD. Unlike by using known peptides from the literature, we try to use as much possible as combination of short peptides. First, we created peptides of random sequences, after that we docked them to spike RBD protein by using AutoDock CrankPep (ADCP). For deciding to use the proper docking tool, we have done a comparison work between free docking tools (Vina and ADCP) which are developed by AutoDock Software. After finishing the dockings, top ones with highest binding energy result are selected for the next step which is Molecular Dynamics (MD) simulations. Results of simulations are controlled according to their RMSD trajectory and binding energy. At the end, one sequence, "wfdwef", stood out as promising over the others. For further, this finding can be used as potential inhibitor for coronavirus after experimental studies
-
ÖgeThermodynamic stability of binary compounds: A comprehensive computational and machine learning approach(Graduate School, 2024-06-06)Exploration and exhaustive comprehension of novel materials are the main objectives of materials science. Laboratory evaluations have been the primary method by which substantial advancements have been achieved throughout the development of this scientific field. The contributions of density functional theory (DFT) algorithms have significantly altered the field of materials science over the past twenty years. These algorithms balance accuracy and efficiency. Supercomputers have enabled substantial breakthroughs in predicting electrical properties of crystal formations, facilitating a fundamental transition in the discipline. Developments of robust algorithms and lower computing costs have made data-driven approaches in materials research more widely adopted. Researchers can now analyze enormous datasets to guide experiments and uncover novel materials. Although databases are frequently used in contemporary materials science, there are some gaps regarding phonon calculations and the thermal properties of compounds. To address this deficiency, this thesis calculates the phonon stability, heat capacities at 298.15 K, formation enthalpies, formation entropies, and Gibbs free energies of binary structures. A total of 879 binary structures were examined, and the results of these calculations were compiled into a data set. In a recent study by my research team, the formation enthalpies and mechanical strengths of binary structures at absolute zero were investigated. This thesis contributes to this work by providing detailed analyses of the dynamic stability and thermodynamic properties of the same binary structures, supporting the findings of my team's prior research. In the initial phase of this thesis, the thermodynamic properties and phonon stabilities of the compounds were calculated. Subsequently, inspired by the PN-PN table model proposed and utilized in our team's recent work, this data set was mapped and visualized on a PN-PN table according to the periodic numbers (PN) assigned to the elements in the structures. This approach enabled the integrated visualization of phonon stability and other thermodynamic properties. Consequently, the chemical similarities between structures were more easily comprehended through the groups in the map, and the so-called forbidden regions were highlighted. Forbidden regions are regions in which specific pairings of elements are unable to form stable phases, which provides critical information on stability based on the PN numbers of the elements. The basic principle of the periodic numbering approach is as follows: First, periodic numbers (PN) are assigned to the elements with respect to their electronegativity, principal quantum number, and valence shell configuration, and then this numbering is extended to binary systems. This makes it easier to understand the chemical trends in the compounds formed by the elements and to predict phase formation. Although there are some exceptions in this mapping, it clearly shows the structures where phase formation is not expected. In our team's previous work, the PN-PN table significantly facilitated the identification of critical regions in different chemical systems and allowed for the analysis of trends in the chemical properties of equiatomic binary phases. Based on this, density functional theory-based thermodynamic calculations were performed in this thesis, providing thermodynamic data supporting the inferences of formation enthalpy and crystal structure stability calculated in our team's previous studies. A total of 879 structures' phonon stabilities were determined, and heat contribution values were calculated. Thus, the phonon stability and heat contribution data obtained from this thesis can be integrated with the mechanical strength properties of the structures from our team's previous findings. This allows for a more detailed interpretation of the relationship between phonon and mechanical stability. Additionally, using the elemental and structural properties of the compounds, machine learning techniques were applied to the current data set. Random Forest, Support Vector Machines (SVM), Gradient Boosting, and Decision Trees were assessed for their capacity to predict phonon stability. The Decision Tree model exhibited the highest performance, with an accuracy rate of 80\%. These models' accuracy was significantly enhanced by elemental descriptors such as band center, mean covalent radius, and mean electronegativity. The band center indicates the effect of the position in the electronic band structure on phonon stability, the mean covalent radius reflects the bonding properties of atoms, and the mean electronegativity determines the atoms' tendencies to attract electrons, thus affecting phonon stability. For predicting Gibbs free energy, Random Forest Regression, K-Nearest Neighbors (KNN) Regression, Support Vector Regression (SVR), and Linear Regression models were used. The performance of these models was evaluated using a 5-fold cross-validation method. The Random Forest Regression model exhibited the highest performance with an average score of 0.846. This result indicates that Random Forest Regression is the most effective model for predicting Gibbs free energy. These findings may encourage the broader application of machine learning techniques in future research. This significant step in understanding and modeling thermodynamic properties plays a critical role in optimizing material structures. In the future, it is expected that the methods of this study will be adapted and developed more specifically for certain material classes or other academic applications. This approach also serves as an efficient example of the discovery and design planning processes in materials science.