Graduate School of Engineering and Science
Permanent URI for this collectionhttps://hdl.handle.net/11693/115678
Browse
Recent Submissions
Item Open Access Generating multiple-choice knowledge questions with interpretable difficulty estimation using knowledge graphs and large language models(2025-10) Şakiroğlu, Mehmet CanGenerating multiple-choice questions (MCQs) with difficulty estimations remains challenging in automated MCQ-generation systems. This thesis proposes a novel methodology for generating MCQs with difficulty estimations from the given text documents by utilizing knowledge graphs (KGs) and large language models (LLMs). Given a set of documents, the approach proposed in this thesis uses an LLM to construct a KG, from which MCQs are then systematically generated. Each MCQ is generated by selecting a node from the KG as the key, sampling a related triple or quintuple—optionally augmented with an extra triple—and prompting an LLM to generate a corresponding stem from these graph components. Distractors, the wrong choices, are then selected from the KG. For each MCQ, nine difficulty signals are computed and combined into a unified difficulty score using a data-driven approach. Experimental results demonstrate that our method generates high-quality MCQs whose difficulty estimations are interpretable and align with human perceptions. Our approach improves MCQ generation by integrating structured knowledge representations with LLMs and a data-driven difficulty estimation model.Item Open Access Sequence-to-graph alignment on a processing-in-memory system(2025-09) Öztürk, Ömer YavuzGenome graphs are used to represent the genetic information and variation of a population rather than a single individual. The sequence-to-graph alignment (SGA) problem can be defined as finding the best match between a query sequence and a genome graph. SGA algorithms are expected to suffer from a memory bottleneck due to the irregularity of graph representation, and benchmarks confirm more than 50% memory boundness found in some applications. Therefore, the SGA alignment problem can benefit from processing-in-memory (PIM) technologies. PIM is an upcoming non-von Neumann architecture that allows computing near the main memory without the need to utilize the memory bus for data transfers to the CPU and back. One of the currently available PIM technologies developed by UPMEM is an architecture consisting of thousands of DPUs (DRAM Processing Units) that allow for much faster and energy-efficient memory access. Our contribution includes SEGAPIM, the design of a pipeline that partitions and distributes a genome graph across DPUs, directs short reads to the relevant DPUs according to their seed locations, and performs alignment solutions within DPUs. Although these distribution and routing components are part of the envisioned full system, the implemented portion of our work focuses on calculating alignment scores for each seed of each read using an adaptation of the graph wavefront alignment algorithm (GWFA) under very limited memory, followed by collecting and finalizing the results on the host CPU. We present comparisons for accuracy and run-time of our implementation to state-of-the-art tools such as vg and GraphAligner, and conclude that the PIM architecture at hand provides a promising speed-up and energy-saving for the SGA problem.Item Embargo Robustness and plasticity in Drosophila embryogenesis: a comparative analysis of nearly isogenic and non-isogenic mass populations(2025-09) Çalıkoğlu, TunahanRobustness and plasticity are fundamental aspects of early development, but how they are shaped by genetic background and environment remains unclear. To address this, we examined the effect of temperature perturbation on the early embryonic markers in Drosophila melanogaster across widely used isogenic laboratory lines with potential genetic differences, natural populations from Saimbeyli/Adana, and weak bcd mutants. Early embryonic markers included the Bcd gradient and Hunchback (hb) and even-skipped (eve) stripe positions. At 25°C, most wild-type isogenic strains had comparable hb pbps on average. On the other hand, Canton S had the most variable (5.2) hb pbp %EL, whereas the Saimbeyli Mass population had the most stable, least variable (1.5) pbps at the same temperature. At 29°C, Oregon R hb pbps shifted posteriorly, while the Saimbeyli Mass line remained unchanged compared to 25oC. Canton S had the highest variation (5.05) Oregon R had the lowest variation (1.4). Eve second stripe was generally more robust against temperature change than hb, though Berlin K, Oregon R, and Saimbeyli Mass showed significant shifts towards posterior. For eve second stripe at 29oC, Saimbeyli 2 had the highest variation (2.8) and Berlin K had the lowest (0.8). In weak bcd mutant lines, steepness of Bcd gradient increased at 22°C, which correlated with the increase in embryo-to-embryo variations. They also showed lower survival rates which correlated with increased mutant cuticle phenotypes. In contrast, wild-type isogenic and Saimbeyli lines maintained high viability under all conditions. Notably, the Saimbeyli mass line showed more robust (robust averages across temperatures) but more variable (increased embryo-to-embryo variation) hb pbps, but greater variability and plasticity in eve expression. These findings indicate that robustness and plasticity are not uniform outcomes of development, but depend on the segmentation network, the developmental marker, and maternal inputs. Development can buffer some traits while allowing flexibility in others, balancing stability in response to the environmental perturbations.Item Embargo A novel multi-sequence optimization framework for nonlinear regression with gradient boosting machines(2025-09) İlhan, EmirhanGradient Boosting Machines (GBMs) consistently achieve state-of-the-art performance on a wide range of machine learning applications, particularly for problems tabular data. However, their underlying optimization mechanism largely relies on a greedy form of functional gradient descent. This classical approach, while effective, can be path-dependent under noise and prone to locally optimal updates that yield globally suboptimal models. To address these fundamental limitations, we propose a multi-sequence framework that integrates principles from modern optimization directly into the gradient boosting process.Unlike prior work that attempts direct adaptations of vector-based optimizers such as Nesterov's accelerated gradient, our method employs a decoupled architecture inspired by stochastic primal averaging (SPA). This architecture provides a stable foundation upon which we build a series of adaptive target updaters that translate the mechanics of modern optimizers such as Adam into principled learning signals for weak learners. We provide a theoretical analysis, including a linear convergence proof for our base model under standard smoothness and strong convexity assumptions. Our experimental results demonstrate significant improvements in stability and accuracy over LightGBM and other baselines, proving our framework's ability to offer a new level of control over the optimization trajectory while preserving the simplicity and modularity that make GBMs practical, leading to more robust and powerful gradient boosting models.Item Embargo Incomplete preferences and robust satisficing in multi-armed bandits(2025-09) Yıldırım, Yaşar CahitThis thesis studies sequential decision-making in multi-armed bandits (MAB) in two settings: when there are multiple objectives to consider and preferences among actions are non-total, and when actions are subject to adversarial perturbations. In the first part, we introduce PaVeBa, a pure-exploration algorithm for vector-valued rewards under incomplete preferences, using cone-induced comparisons instead of scalarizations. We analyze its behavior and validate it empirically against strong baselines. To support research and reproducibility in this area, we also develop VOPy, a Python library that implements PaVeBa and a broader suite of vector-optimization algorithms and tools. In the second part, we investigate robust satisficing under adversarial attacks. Building on existing robust satisficing formulations (e.g., stability radius and fragility-based), we design a Thompsonsampling variant of an algorithm family and demonstrate that it achieves reliable satisficing performance under targeted perturbations.Item Open Access Ultrafast spectroelectrochemistry of mixed ionic–electronic polymer conductors(2025-09) Altuğ, FatmaElectron transfer and switching mechanisms in conducting polymers remain an open question. When conventional electrochemical techniques such as cyclic voltammetry are employed alone, the measured current inevitably contains contributions from both the redox process of the polymer and the capacitive charging current arising from the voltage-dependent capacitance. Thus, isolating the pure redox current requires additional deconvolution methods. Furthermore, the conventional understanding, repeatedly emphasized in the literature, is that electron transfer in conducting polymers proceeds through an ion-coupled mechanism, in which counter-ions from the electrolyte penetrate into the polymer to maintain charge neutrality. In this thesis, we challenge this conventional description by employing operando ultrafast cyclic voltammetry–UV–Vis spectroscopy, in which the potential is rapidly scanned within a defined range while the spectroscopic response is simultaneously recorded. By extending the sweep rate up to 200,000 V/s, we aimed to suppress net counter-ion displacement. Under these conditions, the observed color change demonstrated that electron transfer can occur even in the absence of ion participation. To this end, operando ultrafast cyclic voltammetry–UV–Vis spectroscopy was utilized. The average spectroscopic responses obtained at ultrafast cycles were deconvoluted into their minimal components using principal component analysis (PCA), and the dependence of the resulting coefficients on the sweep rate was systematically examined. Analysis of these coefficients revealed that, with increasing sweep rate, the relative lifetime of the more oxidized and reduced states changes: counter-ion diffusion becomes increasingly limited, making doping less efficient and leading to a greater contribution from the reduced state. Moreover, optical charge values derived from the PCA coefficients approached steady-state at high sweep rates. These findings provide direct spectroscopic evidence that electron transfer in conducting polymers can proceed independently of ion motion under ultrafast potential modulation.Item Embargo Reducing communication volume increase in latency efficient store-and-forward scheme(2025-09) Uzel, Salih DenizSparse matrix operations such as sparse matrix-vector (SpMV) are latency bound applications where number of messages sent by processors dominate the overall communication overhead. In prior work, Store and Forward (STFW) scheme was proposed and implemented to scale such fine-grain operations as SpMV by operating on specialized Virtual Process Topologies (VPTs). However, the STFW scheme incur an increase in the communication volume due to the STFW overhead. This increase in total communication volume disturbs the scalability of the STFW scheme, especially in sparse matrix-matrix (SpMM) kernels where a sparse matrix is post-multiplied by a tall-and-skinny dense matrix. In this work, we propose and implement a Kernighan-Lin-based (KL) heuristic to iteratively improve the one-to-one mapping produced by the initial task partition. The objective of this heuristic is to reduce the increase in the total communication volume to be incurred by the STFW scheme in SpMV, and SpMM type of kernels. We evaluated the effectiveness of the proposed KL-based heuristic in improving the Total Communication Volume and Maximum Weighted Communication Volume metrics through experimental performance comparisons using VPTs of various dimensions and constructive mapping methods.Item Embargo PDMS surface functionalization for sustained cell adhesion in microfluidic organ-on-chip models(2025-09-15) Erdoğan, EcemThe inherent hydrophobicity of Polydimethylsiloxane (PDMS) poses significant challenges for its use in cell culture applications, particularly in organ-on-chip technology. This study presents a novel, versatile, and biocompatible surface treatment method for polydimethylsiloxane (PDMS), designed to significantly improve its cytocompatibility and shear stability in organ-on-chip applications. Four tailored surface functionalization protocols were systematically explored using 3-(Trimethoxysilyl)propyl methacrylate (TMSPMA), optimized to enhance wettability, surface roughness, and chemical functionality, leading to robust and sustained cell adhesion. The efficacy of these treatments was confirmed via comprehensive surface characterizations of contact angle measurements, scanning electron microscopy (SEM), atomic force microscopy (AFM), and X-ray photoelectron spectroscopy (XPS). The changed hydrophilicity supported robust cell adhesion, proliferation and viability that were determined using three distinct human cell types: HUVECs, MCF-7, and Human-derived induced pluripotent stem cells (iPSCs). Importantly, our findings show that the treated PDMS supports long-term adhesion and proliferation under both static and dynamic microfluidic conditions, with minimal cell detachment even under flow rates corresponding to physiological shear stress (~1.5 dyne/cm²). Numerical simulations further validate the experimental flow conditions, strengthening the model's physiological relevance. This study underscores the effectiveness of TMSPMA treatment in improving PDMS compatibility for cell culture, offering an accessible, tunable, and scalable solution for researchers aiming to create more reliable and biologically relevant microfluidic systems.Item Open Access Awe in built environments: how architectural style and viewing perspective influence visual recognition memory(2025-09) Ekinci, Beyza GülzehraIn recent years, there has been a growing interest in neuroarchitecture. This emerging field integrates neuroscience, psychology, and architecture to investigate the built environment’s impact on cognition, emotion, and behavior. Aesthetic and emotional responses to architecture, especially the experience of awe, are becoming more recognized as significant factors influencing cognitive processes, including attention, memory, and decision-making. However, the association between awe, components of architecture, and perceptual perspective is still insufficiently examined. The present study investigated the interactions among awe, architectural style, and viewing angle on visual recognition memory in built environments, employing a 2 (awe: awe-evoking vs. non-awe-evoking) × 2 (style: traditional vs. modern) × 2 (angle: eye-angle vs. low-angle) within-subjects design. Thirty-six participants participated in a visual recognition memory task involving real-life architectural imagery. Analyses of recognition sensitivity d ′ indicated that traditional architecture revealed higher recognition sensitivity than modern architecture, and eye-angle perspectives yielded better sensitivity than low-angle perspectives. Awe alone did not yield a significant main effect; however, it interacted with architectural style, diminishing discriminability for modern architecture while having no effect or a slight improvement on traditional architecture. Analyses of the response criterion c revealed that awe, style, and angle independently modulated decision thresholds, with awe-evoking images eliciting a more liberal bias. This study contributes to the existing literature by establishing the contextual restrictions of awe’s mnemonic effects, indicating that architectural stimuli’s perceptual and stylistic characteristics moderate its influence on memory. The combination of affective science, neuroarchitecture, and signal detection theory provides a new framework for analyzing the interplay between aesthetic and perceptual factors in shaping memory within the built environment.Item Embargo Molecular interaction mechanisms of bulky ions on aqueous neutral macromolecules(2025-09) İşsever, ErtanThe ion specific effects on the solubility of macromolecules have been demonstrated for the first time more than a century ago yet the molecular level detailed picture of ion – macromolecule interactions has not been fully drawn. When considering the effects of salts on macromolecules, it was generally considered that the overall effect is dominated by the anions, in which the counter cations acting mainly for the charge balance in the medium. Recently, the chloride salts of tetra alkyl ammonium (NR4+) ions refuted such assumptions via demonstrating that weakly hydrated cations can interact with macromolecules as effectively as weakly hydrated anions (I-, NO3-, SCN-). In this thesis, we studied how tetra alkyl ammonium cations (NR4+) behave when in conjunction with surface-active anions (I-, NO3-) and whether the ion effects on macromolecules are additive. When we examined the phase transition temperature changes of tetra alkyl ammonium iodide (NMe4I, NEt4I, NPr4I, NBut4I) salts on PNIPAM, a thermo-responsive polymer that enables to construct a thermodynamic model. A surprising salting-out behavior was observed, particularly for tetra butyl ammonium iodide (NBut4I) even though a strong salting-in behavior is expected due to the presence of two surface-active ions. To elucidate the mechanism of this non-additive behavior, we investigated the solubilized and collapsed states of the macromolecules using 1H-NMR and ATR-FTIR experiments, respectively. Our experimental results indicated that when the cation and counter-anion are bound simultaneously, the charge of the macromolecular surface remains close to neutral, and thus, a salting-out behavior can occur. On the other hand, when the hydrophobicity of the cation decreases, this cooperative binding is partially disrupted, and the weakly hydrated anion becomes dominant, and clear salting-in effect is observed. In the second part, tetraphenyl-borate (TPB-) and tetraphenyl-phosphonium (TPP+) ions are also investigated, which are bulky molecular ions with charge asymmetry. Hydration spectroscopy (MCR-Raman) demonstrated the influence of charge asymmetry on the interaction of pi electrons of the phenyl groups and hydration water via the dangling O-H vibrational band shifts. In light of this, the effects of these ions on three different model neutral macromolecules (PNIPAM, PDEA and ELP) were investigated. The anion with the strong water interaction (TPB-) exhibited a salting-out behavior, while the cation with the weakest water interaction (TPP+) exhibited a salting-in action. NMR and ATR-FIR experiments demonstrated that the tetraphenyl phosphonium ion (TPP+), in particular, showed an apparent non-site-specific interaction with the macromolecule. Consequently, the different hydration fingerprints of hydrophobic ions directly correlated with the well -known strongly hydrated, and weakly hydrated ion concepts, and clearly rationalize their effect on neutral macromolecule.Item Open Access Energy-delay and energy-age trade-off in sleep-wake scheduling of queuing systems(2025-09) Gürsoy, ÖmerMinimizing energy consumption while satisfying Quality of Service (QoS) requirements has become a critical objective for the efficient and successful deployment of next-generation networks (NGNs). Sleep-wake scheduling stands out as a prominent approach among various mechanisms that aim to enhance energy efficiency. This mechanism allows the node to save energy by temporarily turning off the components of its transmission module, albeit at the cost of potential QoS degradation. Numerous sleep-wake scheduling strategies have been proposed in both academic and industrial contexts to manage the transition between Sleep and Active operational modes. Most of the existing methods focus on a narrow set of QoS metrics while employing conventional algorithmic approaches. As communication systems grow in complexity and diversity, there is an increasing need for more advanced mechanisms that can accommodate a broader range of QoS considerations. In parallel, the advances in computational capabilities support the integration of sophisticated control systems, such as model predictive control (MPC), which offer greater adaptability to dynamic environments. Motivated by these developments, this thesis explores the application of innovative sleep-wake scheduling strategies in communication systems. In particular, we consider (i) delay (ii) information freshness, as the QoS metrics for the analysis and control of sleep-wake scheduling in two separate settings. In the first setting, we introduce a novel open-loop dynamic coalescing method for Energy Efficient Ethernet (EEE), which leverages queuing theory and is inspired by MPC. Our first proposed method, referred to as MPC-mean, aims to reduce the energy consumption of Ethernet links while maintaining constraints on the average queue waiting time. Addition ally, we propose another MPC inspired approach, called MPC-tail, which focuses on controlling the tail distribution of queue waiting times. Beyond Ethernet interfaces, we investigate information freshness in Internet of Things (IoT) sensor networks, where nodes alternate between Sleep (or Low-power) and Active modes. We analyze the trade-off between energy consumption and the Age of Incorrect Information (AoII) using timer-based sleep-wake scheduling. To validate the performance and efficiency of our proposed techniques, we provide extensive numerical simulations.Item Embargo Ion specificity and cononsolvency effects on neutral macromolecules(2025-09) Özdoğan, Yaren ŞevvalIons, osmolytes, and other small molecules can interact with macromolecules and affect their properties, such as solubility and hydration. The ion specificity has been an important research topic since Franz Hofmeister published the “Hofmeister Series”, which ranks the anions and cations based on their ability to increase or decrease the solubility of macromolecules. It is known that the weakly-hydrated anions can interact with macromolecules, while the strongly hydrated ones are excluded from the macromolecular surface due to having strongly interacting hydration shells. In this thesis, the ion-specific effects are extended to sugar-based macromolecules from commonly studied amide-based ones. Methylcellulose (MC), hydroxypropyl cellulose (HPC), and dextran were utilized as model polymers for investigating whether the Hofmeister anions can interact or not. Lower Critical Solution Temperature (LCST) measurements showed that MC and HPC are affected by the Hofmeister anions with the same way as the amide-based macromolecules. Namely, SCN- act as a surface-active ion, and salts-in whereas SO42- ion salts-out macromolecules via ion exclusion mechanism. The proton nuclear magnetic resonance (1H-NMR) spectroscopy measurements showed that the methyl groups of MC are responsible for the interaction between the Hofmeister anions and the polymer. In the second part, “cononsolvency” which is an important phenomenon that occurs when two good solvents decrease the solubility of the macromolecule when they are mixed. For understanding the mechanism of this phenomenon, poly(N-isopropylacrylamide) (PNIPAM) and poly (N, N - diethylacrylamide) (PDEA) were used as model macromolecules. PNIPAM showed cononsolvency behavior via the addition of methanol, ethanol, and isopropanol, while PDEA did not exhibit the same anomaly, rather showed cosolvency. Attenuated Total Reflection – Fourier Transform Infrared (ATR-FTIR) spectroscopy combined with Multivariate Curve Resolution (MCR) analysis showed that both the changes in the bulk properties of solvent-cosolvent mixtures and the hydration shell of the macromolecules in the collapsed state cause the cononsolvency behavior for PNIPAM. Our results show that the interface and preferential alcohol adsorption dictate the cononsolvency phenomenon.Item Embargo TPTF: leveraging global receptive fields and spectral filters for visual robotic manipulation with transporting transformer networks(2025-09) Şenol, Barış BilginTransformers have recently emerged as a powerful and versatile tool capable of capturing complex interactions among long-distance features, making them highly suitable for learning visual representations for robotic manipulation tasks. However, existing density estimation models, such as Transporter networks [1], rely on convolutional backbones that primarily process local information, requiring multiple convolutional stems to learn task-specific policies. We explore the potential of Transformer networks and alternative token mixing mechanisms for categorical density estimation and propose the Transporting Transformer networks for complex robotic pick-and-place tasks. Our approach employs a single encoder stem to leverage global features for learning both pick and pick-conditioned place policies. Building upon its enhanced capacity, we also introduce a novel training scheme for multi-task learning on the Ravens benchmark. The Transporting Transformer learns manipulation policies directly from visual observations without object-level assumptions, achieving improved performance through effective modeling of long-range spatial relationships. It maintains sample efficiency comparable to existing methods while demonstrating superior performance in both single-task and multi-task learning settings.Item Open Access Spectrometer frameworks for gradient arrays(2025-09) Öztürk, Mehmet EminGradient arrays are gaining popularity in the magnetic resonance imaging (MRI) community due to their superior performance and flexibility compared to conventional gradient coils. The technology can be used to excite multiple slices simultaneously, mitigate the B1+ inhomogeneity, reduce the power loss on the cryostat, and increase peripheral nerve stimulation (PNS) thresholds. However, despite their potential, fully standalone MRI systems based on gradient arrays are yet to be realized. With an increasing number of channels, scalable design methodologies become essential to resolve the hardware complexity. In this thesis, we present the development of modular spectrometer architectures for gradient array coils and parallel radio frequency (RF) transmit arrays to work towards a fully standalone gradient array magnetic resonance imaging system. Building upon previous work, the spectrometer submodules are implemented entirely on field programmable gate array (FPGA) hardware and consist of three main subsystems: a sequence parser, a gradient controller, and an RF controller. The sequence parser reads PulSeq-formatted conventional sequences and converts them into hardware-compatible memory structures. The gradient controller accesses these memory-mapped sequences and generates time-resolved gradient amplitudes, which are projected onto array channels and sent to an external driver chain. Lastly, the RF controller synthesizes a carrier clock and drives external RF power amplifiers using pulse-width-modulated (PWM) envelope signals. Unlike traditional RF synthesizers that rely on digital-to-analog converters (DACs) or mixers, the RF clock is generated entirely using FPGA primitives, simplifying the overall architecture and reducing scaling costs.Item Embargo Implications of the anisotropic mexican-hat band structure in 2D materials on their thermoelectric properties: an analytical and computational investigation(2025-09) Hilal, MuhammadThe electronic band structures of two-dimensional (2D) materials often exhibit non-parabolic, highly anisotropic features near the valence band edge. A particularly intriguing case is the Mexican-hat-like dispersion, characterized by a ring-shaped energy extremum in momentum space. This topology gives rise to a van Hove singularity in the 2D density of states (DOS) and a sharp onset in the electronic transmission spectrum, both of which can significantly enhance thermoelectric performance. However, realistic materials frequently exhibit angular anisotropy in these dispersions, shifting the DOS singularity and degrading transport efficiency. In this thesis, we investigate the interplay between anisotropy in Mexican-hat-like band structures and thermoelectric properties across a family of 2D materials. Using first-principles density functional theory (DFT) and density functional perturbation theory (DFPT), we systematically study seven monolayer compounds, including PbBrF, PbClF, PbIF, BaHI, BiOCl, CaHBr, and SrHI. For each material, we perform structural optimization, electronic band structure analysis, and phonon stability checks. Thermoelectric transport coefficients are computed using the semi-classical Boltzmann transport equation via BoltzTraP, incorporating spin-orbit coupling (SOC) for the PbXF compounds. Phonon-limited lattice thermal conductivity is evaluated through third-order force constants obtained from thirdorder.py and ShengBTE. To analytically interpret the Mexican-hat features, we develop a tight-binding model parameterized by the ratio of next-nearest to nearest-neighbor hopping amplitudes (ξ = t2/t1) and an angular anisotropy term β. We demonstrate how ξ and β jointly control the curvature and shape of the band edge, thus tuning the thermoelectric response. Across the PbXF series, SOC suppresses the valence-edge singularity and tends to lower the p-type performance while modestly enhancing the n-type response by sharpening conduction edges. Phonon calculations (second- and third-order force constants; ShengBTE) reveal intrinsically low lattice thermal conductivities due to soft optical modes and strong three-phonon scattering. As a case study in band-structure engineering, we apply biaxial tensile strain (0–6%) to SrHI and show that strain reduces the Mexican-hat anisotropy, slightly decreases the hat height and ring radius, and sharpens the DOS onset. Consistent with the Mott relation, p-type S and the power factor improve (with a mild trade-off in conductivity). In contrast, the lattice thermal conductivity decreases as strain softens modes and enlarges anharmonic phase space. Taken together, the results establish clear descriptors and knobs—ξ, β, hat height, and strain—for designing high-performance 2D thermoelectrics.Item Open Access Direct volume rendering of tree-based tetrahedral adaptive mesh refinement data(2025-09) Ünalan, Musa EgeRay-tracing-based direct volume rendering (DVR) techniques often use data representations such as regular grids, unstructured meshes, and adaptive mesh refinement (AMR) data. A less-explored option for DVR is tree-based tetrahedral AMR data (Tet-AMR), which combines the benefits of unstructured meshes and AMR by having both a coarse unstructured tetrahedral mesh that can represent complex domains, on which acceleration structures can be constructed to perform efficient ray-triangle intersection tests, and a forest of refinement trees, each rooted at a coarse mesh element, increasing detail where needed. Tet-AMR data can be visualized by converting it to an unstructured mesh representation; however, this approach introduces new unstructured elements, increasing memory usage during rendering while decreasing the performance of acceleration structures. We propose leveraging the regularly subdivided nature of the tetrahedral refinement trees by only storing the coarse geometry during rendering. We construct a bounding volume hierarchy over the coarse mesh to efficiently identify the refinement trees from which to sample. Then, we generate the geometry of the finer level elements on the fly when traversing the refinement trees to find the actual elements to sample. We also show that the tree structure can be utilized to implement a dynamic view-dependent level of detail effect, improving performance by decreasing fidelity in regions that minimally affect the final image. It can be used to obtain a density range for each tree, enabling empty space skipping with ray marching renderers or used as local majorant extinctions with delta tracking renderers to improve performance. Our proposed method outperforms or performs comparably to rendering the Tet-AMR as an unstructured mesh, using less memory and enabling the described effect and optimizations on the various datasets we have tested with our GPUrenderers.Item Embargo Human-guided subgoal learning for sequential manipulation in narrow and cluttered 2D maze-like environments(2025-09) Haliloğlu, Dilruba SultanRobotic agents should be able to perform sequential manipulation tasks since many real-life tasks consist of interdependent sequential actions. Sequential manipulation in cluttered narrow spaces is a challenging issue in robotic planning because of the high dimensionality of the solution space and existence of local minima. One approach to solve such complex sequential manipulation tasks is to develop algorithms capable of decomposing these tasks into manageable subproblems. This work presents a framework that learns critical subgoals in a scene configuration from human data in 2D maze-like environments involving sequential object manipulation. We designed a data collection interface that allows participants to annotate subgoals for agents and objects in narrow and cluttered environments. We then use the collected data to train deep neural networks that predict subgoal target entity and possible subgoal positions as distributions in a given configuration. Finally, subgoals sampled from the predicted distributions are used to construct a final sequential manipulation plan to solve the task. The proposed pipeline generates a sequential manipulation plan for a diverse set of tasks with a success rate of 95.9%, demonstrating that using human-guided critical subgoal generation is a viable and promising approach.Item Embargo Contact-VLA: zero-shot planning and control for contact-rich manipulation(2025-09) Çiçek, BerkVision-Language-Action (VLA) systems often lack adaptability and explainability due to their blackbox structure and dependency on fixed action sets from extensive tele-operated datasets, limiting their effectiveness in complex, dynamic manipulation scenarios. To address this issue, we propose a novel VLA framework capable of effectively managing complex, dynamic, and contact-rich manipulation tasks. By integrating foundational vision-language models with motion planning and reactive controllers, our system achieves zero-shot planning and adaptive manipulation without relying on extensive tele-operated action datasets. Unlike conventional VLAs, we explicitly separate the roles of Vision-Language Models (VLM) and Large Language Models (LLM): the VLM handles object parameter extraction and environmental modeling, while the LLM generates initial contact strategies and cost estimations. These two components collaboratively contribute to the creation of a simulated environment in which our dynamic planner operates. Additionally, this modular approach significantly enhances both the explainability and performance of the overall framework, as demonstrated through our rigorous ablation studies. Furthermore, we introduce a memory unit to leverage past manipulation experiences, enabling the generalization and efficient reuse of learned contact strategies and parameter adjustments across diverse manipulation scenarios. Experiments conducted on challenging contact-rich tasks validate our framework’s robustness and highlight the critical design elements that contribute to its effectiveness.Item Open Access Normalized least mean square based optimization algorithms for deep learning and their applications(2025-08) Türeyen, EsraAdaptive filters are online learning systems that adjust their parameters to achieve a desired output via an optimization algorithm. They are widely used in signal processing applications like noise cancellation, echo suppression, channel equalization, and system identification. An algorithm used to adjust the parameters of adaptive filters is Least Mean Square (LMS) which minimizes mean square error of the cost function. Normalized Least Mean Square (NLMS) algorithm is a refined version of LMS, which aims to improve stability and convergence by scaling the learning rate based on the power of the input vector. Adaptive filters lay the foundation of many learning systems, and their update rules resemble gradient-based learning found in modern machine learning. Input Normalized Stochastic Gradient Descent (INSGD) optimizer is recently introduced as a variant of Stochastic Gradient Descent (SGD), and utilizes the input normalization approach in NLMS to improve convergence behavior. Along with convergence issue, deep learning optimization has many concerns such as accuracy, stability, adaptability, vanishing gradients, sensitivity to hyperparameter tuning. Various optimizers are proposed in literature with different approaches that aim to address any of the purposes or problems. As an extension to current literature, this thesis broadens the application scope of INSGD optimizer by applying it to language tasks. Specifically, INSGD is used to train the Bidirectional Encoder Representations from Transformers (BERT) model on the General Language Understanding Evaluation (GLUE) benchmark. Demonstrating improved performance, INSGD emerges as a promising alternative to standard SGD in Natural Language Processing (NLP) tasks. Building on this idea, a novel optimizer named Input Normalized Adam (INAdam) is proposed. INAdam integrates the input normalization principle derived from the NLMS algorithm into the Adam optimization framework. The effectiveness of INAdam is shown through extensive image classification experiments on CIFAR-10 and CIFAR-100 datasets using the Residual Network (ResNet) architecture. As a final contribution, a new optimizer called Loss and Input Normalized Stochastic Gradient Descent (LINSGD) is introduced. LINSGD combines both error normalization and input normalization for dynamic learning rate scaling. The performance of LINSGD is evaluated through classification experiments on CIFAR-10 and CIFAR-100 using ResNet models, highlighting stability characteristics of the proposed optimizer.Item Embargo Repurposing trifluoperazine in combination with kinase inhibitors for glioblastoma therapy: insights from in vitro and larval zebrafish in vivo models(2025-08) Acar, RanaGliomas are classified into lower-grade gliomas and glioblastomas, with the latter representing the most aggressive and treatment-resistant form, highlighting the urgent need for more effective therapeutic strategies. Recently, the antipsychotic drug trifluoperazine emerged as a potent anticancer drug for multiple cancers, including gliomas. Our previous research identified a potent therapeutic combination of trifluoperazine with sorafenib, a well-established multi-kinase inhibitor with proven efficacy in hepatocellular carcinoma, in Hep3B cell line model. In the present study, the transcriptomic response to trifluoperazine alone or in combination with sorafenib was evaluated in U87-MG glioblastoma cells and revealed that the combinatorial treatment induced a synergistic effect characterized by growth arrest, reduced invasiveness, and transcriptional signatures of metabolic stress and cell fate reprogramming. While trifluoperazine alone promoted a proliferative gene expression profile through the upregulation of essential metabolic and cell cycle-associated genes, its combination with sorafenib counteracted these effects by suppressing oncogenic signaling and amplifying tumor-suppressive pathways. These findings highlight the potential of trifluoperazine as a repurposed agent in kinase-targeted glioblastoma therapy and underscore the benefit of combinatorial strategies to overcome adaptive resistance mechanisms. Although sorafenib treatment in glioblastoma cells provided mechanistic insights into its synergy with trifluoperazine, in vivo exposure to sorafenib caused severe morphological defects in zebrafish larvae over a 72-hour window from 2 to 5 dpf. Moreover, in a previous glial reporter imaging screen using Tg(gfap:GFP) line zebrafish larvae, sorafenib was observed to markedly reduce GFP fluorescence, whereas trifluoperazine alone did not produce any detectable adverse effects in either assay. This in vivo toxicity prompted the search for alternative targeted kinase inhibitors with more favourable safety profiles to combine with trifluoperazine. Other kinase inhibitors that had a growth inhibitory effect on U87-MG as well as A172 cells were identified using MTT assays. A screen of 157 kinase inhibitors resulted in 12 kinase inhibitors with IC50 values less than 5 μM for both cell lines. Among the most promising with full cytotoxic effect were volasertib, INK128, CAY10626, tamatinib, Torin1, bisindolylmaleimide IX, AZD7762, NH125, and BMS345541. Trifluoperazine was also combined with selected kinase inhibitors, and volasertib and INK128 were identified to elicit significant synergism. On the other hand, several kinase inhibitors with significant inhibitory effects alone exhibited significantly reversed impact by the addition of trifluoperazine. The PLK1 inhibitor volasertib was top prioritized due to its highest synergy score in combination with trifluoperazine, and RNA‑seq was performed on U87‑MG cells treated with 2.5 μM volasertib to investigate its translational relevance. The resulting drug‑induced gene signature was evaluated over the TCGA‑GBM cohort, where high volasertib signature scores correlated with significantly improved overall survival, thereby reinforcing the rationale for its repurposing approach in combination with trifluoperazine for glioblastoma therapy. Finally, the early larval toxicity profiles of prioritized single or combinatorial drug treatments were assessed in vivo, and neither induced developmental abnormalities or phenotypic signs of adverse effects. In this context, LDexplore, an R Shiny application was developed to enable multidimensional analysis of zebrafish larval locomotor behaviour under dark:light alternation, incorporating heatmaps and statistical analyses of velocity and acceleration in response to startle stimuli. The app was used to test whether the zebrafish larvae exposed to trifluoperazine alone or in combination with volasertib exhibited different locomotory behaviour and it was found that although trifluoperazine induced a heightened swimming pattern upon stimulus at higher concentrations, the combinatorial treatment was not significantly disruptive to light induced startle response. These findings indicate that the combination of trifluoperazine and volasertib exhibits strong therapeutic potential, with in vivo applicability supported by larval zebrafish assays. Additionally, other kinase inhibitors demonstrating lower toxicity and significant synergistic effects on glioblastoma cell growth were identified, providing insights for further investigation.