|
BBL 596E/696E Seminars
BLU 596E Seminars
CRN 23045/25213
CRN 24461
|
|
1
BBL 596E
|
| Title |
Challenges and Approaches for Clustering Large-Scale Molecular Datasets Using Similarity-Aware Methods |
| Speaker |
Arshia Lamei |
| Date |
May 7th, 2025 |
| Time |
13:30 (GMT+3) |
| Advisor |
Sefer Baday |
Abstract
Clustering large molecular datasets, often represented by high-dimensional fingerprints, is essential in cheminformatics but poses significant computational challenges. Standard algorithms like Agglomerative Clustering can struggle with scalability (O(N2) complexity) and memory requirements when applied to hundreds of thousands of molecules. This paper reviews these challenges based on preliminary experiments using Morgan fingerprints and Jaccard distance. Advanced generative models are subsequently explored, specifically the regularized transformer approach proposed by Tibo et al. (2024), which aligns generation negative log-likelihood (NLL) with molecular similar ity (Tanimoto). While originally designed for local chemical space exploration, the potential for such similarity-aware generation to facilitate more effective downstream clustering by providing better-structured molecular sets or similarity-informed features is discussed. This perspective is compared with traditional methods like the Butina algorithm, and future directions involving scalable computing and deep learning for tackling large-scale molecular clustering are outlined.
|
| Location |
Informatics Institute
Room 412
|
|
|
2
BBL 596E
|
| Title |
Analysis of Microscopic Data on the Behavior of Metallic Nickel Ore During Magnetic Enrichment Using SAM2 and Tensor Decomposition-Based Zero-Shot Image Segmentation |
| Speaker |
Berk Münci İnanç |
| Date |
May 7th, 2025 |
| Time |
13:30 (GMT+3) |
| Advisor |
Behçet Uğur Töreyin |
Abstract
The accurate separation and classification of minerals are crucial for integrating deep learning-based computer vision approaches into mineral enrichment processes. In this work, labeled microscopic mineral images were processed using various tensor decomposition algorithms and then segmented with the SAM2 model using a zero-shot prediction. The main goal of this study is to understand how different tensor decomposition schemes influence segmentation performance and to support further research in mineral enrichment. The results shed light on the impact of these algorithms on segmentation accuracy, providing useful guidance for choosing the most effective methods in the related field.
|
| Location |
Informatics Institute
Room 412
|
|
|
3
BBL 596E
|
| Title |
Underwater Turbid River Image Restoration Using Diffusion Models |
| Speaker |
Fatima Iqbal |
| Date |
May 7th, 2025 |
| Time |
13:30 (GMT+3) |
| Advisor |
Behçet Uğur Töreyin |
Abstract
Underwater images often face challenges due to turbidity caused by suspended particles, leading to hazy and distorted visuals. The lack of real-life underwater data also reduces the efficiency of trained models. This paper introduces a diffusion models-based denoising architecture to restore under water turbid images. The method first quantifies the turbidity noise by optimizing the variance parameter using a data set that replicates the diffusion process of forward noise addition. The trained U-Net architecture then iteratively reconstructs turbid images by implementing a reverse Markov diffusion chain process. In addition to visual enhancements, restored images are evaluated using perceptual evaluation measures such as entropy and NIQE. The results of this project can contribute significantly to the monitoring of marine ecosystems and the study of fish migration patterns.
|
| Location |
Informatics Institute
Room 412
|
|
|
4
BBL 596E
|
| Title |
Network Attack Classification and Incident Response Support Using a Fine-tuned Large Language Model |
| Speaker |
Haluk Çağatay Kır |
| Date |
May 7th, 2025 |
| Time |
13:30 (GMT+3) |
| Advisor |
Enver Özdemir |
Abstract
The limitations of traditional, manual threat hunting in Incident Response highlight the need for more innovative cybersecurity strategies. Manually crafted rules to detect malicious behavior are not scalable and often yield low accuracy. This research proposes a shift toward integrating Large Language Models (LLMs) into Incident Response processes to enhance threat detection accuracy, lower false positives, and optimize resource allocation. By leveraging LLMs, the project seeks to reduce repetitive tasks, speeding up decision-making and overall incident response effectiveness. The methodology involves fine-tuning the Gemma-2b model on the CIC-IDS2017 dataset and creating a Retrieval Augmented Generation (RAG) agent that offers a natural language interface for log analysis. Results demonstrate that Gemma-2b acts as a conversational AI assistant, improving the efficiency of cybersecurity professionals by automating routine tasks. With an overall accuracy of 96.85%, the model shows strong predictive capabilities in identifying diverse network attacks, underscoring its potential for threat detection and classification in cybersecurity operations.
|
| Location |
Informatics Institute
Room 412
|
|
|
5
BBL 596E
|
| Title |
Weakly Supervised HER2 Equivocal WSI Classification |
| Speaker |
Mücahit Ertano |
| Date |
May 14th, 2025 |
| Time |
13:30 (GMT+3) |
| Advisor |
Abdülkerim Çapar |
Abstract
HER2/neu (ERBB2) gene amplification and over-expression are critical biomarkers in breast cancer, with imunohistochemistry (IHC) used to classify HER2 expression levels. Equivocal (2+) cases present challenges, often requiring additional In Situ Hybridization (ISH) tests, which are costly and time-intensive, to guide treatment decisions. This study developed an automated approach using weakly supervised deep learning methods to classify HER2 IHC equivocal (2 +) whole slide images (WSIs) as HER2 positive or HER2-negative. Tumor samples from 113 breast cancer patients were digitized, annotated according to the intensity of HER2 staining, and analyzed using multiple instance learning. These findings demonstrate the potential of multiple instance learning to streamline HER2 testing, reduce reliance on ISH, and improve decision-making for equivocal cases.
|
| Location |
Informatics Institute
Room 412
|
|
|
6
BBL 596E
|
| Title |
A Comprehensive Survey on Multi-Modal Attention Mechanisms in Vision-Language Models |
| Speaker |
Cansu Nur Demirkıran |
| Date |
May 14th, 2025 |
| Time |
13:30 (GMT+3) |
| Advisor |
Ertuğrul Karaçuha |
Abstract
Recent advancements in Vision-Language Models (VLMs) have led to significant improvements in tasks requiring joint reasoning over visual and textual modalities. A core component driving this progress is the development of sophisticated multi-modal attention mechanisms. This paper provides a comprehensive survey of multi-modal attention architectures in VLMs, including cross-attention, self-attention, co-attention, and hierarchical attention mechanisms. These mechanisms are categorized and analyzed based on their design and effectiveness in benchmark tasks such as Visual Question Answering (VQA), image captioning, and cross-modal retrieval. In addition, computational trade-offs are examined, and emerging challenges and research directions in multi-modal attention are highlighted.
|
| Location |
Informatics Institute
Room 412
|
|
|
7
BBL 596E
|
| Title |
Predicting Hit Songs with Machine Learning |
| Speaker |
Çağatay Kurt |
| Date |
May 14th, 2025 |
| Time |
13:30 (GMT+3) |
| Advisor |
Şule Öğüdücü |
Abstract
Hit song prediction is a subfield of Music Information Retrieval (MIR) that is focused on analyzing hit songs that have appeared in one or more music charts around the world. These charts and their rankings are usually based on physical/digital sales and downloads, streaming/airplay count from radio and web platforms. Excluding the artist based factors, analyzing hit songs by their raw waveforms may unveil highly valuable information. This information can be used by producers and record labels in order to determine which songs are likely to be a hit prior to their release and to discover new talents. We gathered a dataset that contains hit songs from Billboard year-end charts of last 20 years, and equal amount of non-hit songs from same genres. Their high level audio features are extracted from Spotify API and this dataset is used to train five machine learning classifiers (logistic regression, support vector machine, decision tree, random forest, XGBoost) as baseline models. Some of these models were able to achieve 86% accuracy. We proposed new audio-based features as a part of ongoing work.
|
| Location |
Informatics Institute
Room 412
|
|
|
8
BBL 596E
|
| Title |
Medical Image Analysis Using Deep Learning |
| Speaker |
Yusuf Emir Meşe |
| Date |
May 14th, 2025 |
| Time |
13:30 (GMT+3) |
| Advisor |
Abdulkerim Çapar |
Abstract
Deep learning has revolutionized medical image analysis by enabling automated and accurate diagnosis from medical scans such as X-rays, MRIs, and CT scans. Convolutional neural networks (CNNs) and transformer-based models have significantly improved image classification, segmentation, and anomaly detection. These techniques enhance diagnostic efficiency, reduce human error, and assist healthcare professionals in early disease detection.
|
| Location |
Informatics Institute
Room 412
|
|
|
9
BBL 596E
|
| Title |
Interpreting Noise in Pulse-Controlled Quantum Gate Operations: A Comparative Study of Modeling Approaches |
| Speaker |
Negar Dokhtmirzahasanvahid |
| Date |
May 21st, 2025 |
| Time |
13:30 (GMT+3) |
| Advisor |
Deniz Turkpence |
Abstract
Accurately modeling quantum gate operations in the presence of noise is crucial for improving quantum hardware performance and error mitigation strategies. In this work, we investigate the effects of noise on single- and two-qubit gates implemented using pulse control. We compare three different approaches: unitary models, which assume ideal closed-system evolution; nonunitary models, which incorporate decoherence and dissipation; and pulse-based simulations using ModelingLib. By benchmarking these models against experimental results from noisy quantum processors, we analyze their ability to capture real-world noise sources, including decoherence, calibration errors, and crosstalk. Our findings provide insights into the strengths and limitations of each modeling approach, highlighting discrepancies between theoretical predictions and hardware behavior. This work contributes to a more accurate interpretation of noise in pulse-based quantum operations and informs the development of improved simulation techniques for near-term quantum devices.
|
| Location |
Informatics Institute
Room 412
|
|
|
10
BBL 596E
|
| Title |
MaskDINO Instance Segmentation for Chromosome Images: Enhancing Accuracy through Boosting |
| Speaker |
Ikra Nergiz |
| Date |
May 21st, 2025 |
| Time |
13:30 (GMT+3) |
| Advisor |
Behçet Uğur Töreyin
Abdulkerim Çapar |
Abstract
Cytogenetic analysis plays a crucial role in the diagnosis of genetic diseases and involves the examination of chromosomes. Metaphase images are analyzed to detect any numerical or structural anomalies in the chromosomes. This process is error-prone and time-consuming due to human intervention. Errors in this process can lead to inappropriate treatments and incorrect clinical diagnoses. In recent years, deep learning techniques in computer vision have been widely used to automate chromosome analysis and improve the accuracy of chromosome segmentation. This paper demonstrates the effectiveness of MaskDINO for instance segmentation, based on Detectron2 and trained on chromosome image datasets. A boosting technique was applied using custom datasets comprising over 50,000 chromosome images. The model was trained on a specific dataset and then evaluated on a separate one. The 1,000 images with the lowest mAP50@95 scores were added to the training dataset, and the model was retrained on this extended dataset. The scores of the model trained on the extended dataset show a %X improvement compared to the initial training dataset.
Additionally, data augmentation strategies to enhance the model's performance on low-resolution and blurry images were examined.
|
| Location |
Informatics Institute
Room 412
|
|
|
11
BBL 596E
|
| Title |
Review and Implementation of The Article of Dynamic Routing Scheme for Linking Wireless Sensor Network towards Internet of Things |
| Speaker |
İsmail Fırat Sürkit |
| Date |
May 21st, 2025 |
| Time |
13:30 (GMT+3) |
| Advisor |
Ömer Melih Gül |
Abstract
The rapid advancements in wireless sensor networks (WSNs) have paved the way for significant progress in the Internet of Things (IoT), enabling various large-scale applications. For these applications to function effectively, it is crucial that routing protocols ensure efficient and reliable packet delivery to their intended destinations. Although existing routing protocols are designed to fulfill basic network requirements, they often fall short in achieving a balance between energy efficiency and secure data transmission. This imbalance can result in reduced network performance and reliability, particularly in large-scale IoT systems. To address these limitations, we propose a secure and energy-efficient routing scheme specifically tailored to meet the unique challenges of IoT networks. The scheme incorporates a controller node that plays a central role in monitoring the behavior of nodes along the routing path. This includes evaluating node cooperation, energy consumption, and identifying any malicious behavior that could compromise the network. The controller node actively observes data transmission processes, detecting potential issues such as packet loss, transmission failures, congestion, or malicious activity. Upon identifying these challenges, the controller node works collaboratively with neighboring nodes to restore reliable communication by dynamically establishing alternative multipath routes. By integrating these mechanisms, the proposed scheme not only ensures energy efficiency but also enhances the security and reliability of data transmission in IoT networks, making it a robust solution for large-scale deployments.
|
| Location |
Informatics Institute
Room 412
|
|
|
12
BBL 596E
|
| Title |
Machine-learning Based Detection of Mononuclear Inflammatory Cells in the Kidney |
| Speaker |
İrem Ayşe Kayacıoğlu |
| Date |
May 21st, 2025 |
| Time |
13:30 (GMT+3) |
| Advisor |
Volkan Müjdat Tiryaki |
Abstract
Banff Lesion Scores are among the most common systems for deciding whether an organ transplant was successful or not. However, the reproducibility of Banff-scoring is mediocre at best and it is time-consuming in daily practice. Human error exists and differences ensue from one pathologist to another. The aim is to reduce the workload of pathologists and increase the scoring consistency by developing an automated assessment system
|
| Location |
Informatics Institute
Room 412
|
|
|
13
BLU 596E
|
| Title |
Addressing the Leakage Problem Between Virtual Machines in Cloud Environments |
| Speaker |
Berkay Akar |
| Date |
May 28th, 2025 |
| Time |
13:30 (GMT+3) |
| Advisor |
Mehmet Tahir Sandıkkaya |
Abstract
Cloud computing systems, offering increased resource sharing and availability, represent a ubiquitous technology providing businesses and individuals with flexibility and efficiency. However, the leakage problem between virtual machines (VMs) in these systems poses a significant security concern. In the cloud environment, where virtual machines run on the same physical hardware and utilize shared resources, improper configuration of one VM or a successful attack by a malicious actor gaining unauthorized access from one VM to another can lead to leakage of sensitive data or denial of service. This issue underscores the need for cloud providers and users to continually enhance security measures and improve virtual machine isolation. This study examines the leakage problem between virtual machines in cloud environments, emphasizing the need to assess existing security measures and develop new strategies to address this pressing concern.
|
| Location |
Informatics Institute
Room 412
|
|
|
14
BLU 596E
|
| Title |
Machine Learning Guided Ultra-Large Library Screening for Androgen Receptor |
| Speaker |
Doruk Can Ul |
| Date |
May 28th, 2025 |
| Time |
13:30 (GMT+3) |
| Advisor |
Sefer Baday |
Abstract
The androgen receptor (AR) plays a crucial role in various physiological processes and is a key target in treating conditions such as prostate cancer. Identifying potent AR inhibitors from ultra-large chemical libraries presents a significant challenge due to the immense computational cost of traditional docking methods. This study leverages machine learning-driven deep docking to efficiently screen potential AR inhibitors. Initially, deep docking models are trained on a dataset of 1,440 ligands with known binding scores to AR. These trained models are then applied to the extensive ZINC20 database, consisting of approximately 750 million compounds. By prioritizing high-affinity candidates, this approach significantly reduces the number of compounds requiring full docking simulations, thereby improving computational efficiency while maintaining screening accuracy. The proposed methodology enables a more targeted and cost-effective drug discovery pipeline for AR inhibitors, accelerating the identification of promising therapeutic candidates.
|
| Location |
Informatics Institute
Room 412
|
|
|
15
BLU 596E
|
| Title |
Classification of Melanoma Skin Cancer Using Deep Learning-Based Analysis |
| Speaker |
Furkan Yaşar |
| Date |
May 28th, 2025 |
| Time |
13:30 (GMT+3) |
| Advisor |
Volkan Müjdat Tiryaki |
Abstract
Since melanomas are among the most severe forms of skin cancer, prompt and precise diagnosis is essential to successful treatment. In this work, we provide a deep learning-based method for dermoscopic image-based melanoma classification. To automate and enhance diagnosis precision, a convolutional neural network (CNN) model is trained on a dataset comprising benign and malignant skin lesions. Key performance indicators including accuracy, sensitivity, and specificity are used to assess the model. Our findings show that deep learning methods can distinguish between benign and malignant instances with high accuracy, which may help dermatologists detect and diagnose melanoma in its early stages.
|
| Location |
Informatics Institute
Room 412
|
|
|
16
BLU 596E
|
| Title |
UAV-Driven Energy-Aware Data Collection in Robotic Wireless Sensor Networks |
| Speaker |
Beyza Duran |
| Date |
May 28th, 2025 |
| Time |
13:30 (GMT+3) |
| Advisor |
Ömer Melih Gül |
Abstract
Minimizing energy consumption is essential for robotic wireless sensor networks (RWSNs) to ensure long-term operation and efficiency. Traditional static sink-based cluster protocols often increase energy costs for distant nodes, reducing overall network performance. To overcome this, UAV-assisted data collection offers an effective alternative by enabling mobile sink functionality, thus lowering transmission distances and energy usage. However, many approaches overlook the UAV’s limited battery capacity. This study proposes an energy-aware data collection strategy where a UAV selectively visits cluster head (CH) robots based on remaining battery and CH positions. Unvisited CHs transmit their data through multi-hop communication. The proposed strategy achieves zero total common cost in some scenarios and significantly outperforms existing methods in computation time, offering a practical and scalable solution for real-world RWSNs.
|
| Location |
Informatics Institute
Room 412
|
|
|
17
BBL 696E
|
| Title |
Experimental Assessment of Misalignment Effects in Terahertz Communications |
| Speaker |
Hasan Nayir |
| Date |
May 28th, 2025 |
| Time |
15:30 (GMT+3) |
| Advisor |
Ömer Melih Gül |
Abstract
Terahertz (THz) frequencies play a crucial role in the advancement of next-generation wireless systems, primarily owing to their substantial available bandwidths. The inherent limitation of limited range, attributed to high attenuation in these frequencies, can be effectively addressed by implementing densely deployed heterogeneous networks, complemented by Unmanned Aerial Vehicles (UAVs) within a three-dimensional hyperspace. Yet, the success of THz communications relies on the precise alignment of beams. Inadequate handling of beam alignment can lead to diminished signal strength at the receiver, significantly affecting THz signals more than their conventional counterparts. This research underscores the paramount importance of meticulous alignment in THz communication systems. The profound impact of proper alignment is substantiated through comprehensive measurements conducted using a state-of-the-art measurement setup, facilitating accurate data collection across the 240 GHz to 300 GHz spectrum. These measurements encompass varying angles and distances within an anechoic chamber to eliminate reflections. Through a meticulous analysis of the channel frequency and impulse responses derived from these extensive measurements, this study pioneers quantifiable results, providing an assessment of the effects of beam misalignment in THz frequencies.
|
| Location |
Informatics Institute
Room 412
|
|
|
18
BBL 696E
|
| Title |
Interpreting Transformer Models: Understanding Attention Mechanisms |
| Speaker |
Ruken Missonnier |
| Date |
May 28th, 2025 |
| Time |
15:30 (GMT+3) |
| Advisor |
Sefer Baday |
Abstract
This study delves into the inner workings of transformer-based models by analyzing attention patterns. It aims to provide insights into how these models capture contextual relationships and discusses methods for improving model interpretability.
|
| Location |
Informatics Institute
Room 412
|
|
|
19
BBL 696E
|
| Title |
Enabling Memory-Efficient Inference for Large Language Models with Highly Multithreaded, Distributed Architectures |
| Speaker |
Cihan Çalışır |
| Date |
May 28th, 2025 |
| Time |
15:30 (GMT+3) |
| Advisor |
Ayse Yılmazer Metin |
Abstract
The increasing adoption of Large Language Models (LLMs) is hindered by their high GPU memory demands and computational constraints, particularly on resource-limited hardware. This thesis introduces a novel Ray-based distributed architecture designed to optimize low-GPU-memory inference for open-source LLMs. The proposed system integrates advanced optimization techniques, including dynamic model sharding, intelligent KV cache management, and strategies to maximize computational throughput. Specifically, the architecture employs dynamic layer partitioning across GPUs to minimize per-device GPU memory usage and incorporates methods such as Grouped-Query Attention (GQA), expert-aware KV cache routing for sparse models, and mixed-precision inference to enhance memory efficiency and computational performance. The framework is evaluated using representative open-source LLM architectures, demonstrating its ability to enable efficient inference on constrained hardware while maintaining competitive performance against state-of-the-art solutions. This research aims to provide a scalable and accessible approach to low-GPU-memory LLM inference, addressing a key challenge in broadening access to advanced language models.
|
| Location |
Informatics Institute
Room 412
|
|
|
20
BBL 696E
|
| Title |
Tensor Decomposition-Based Methods in Hyperspectral Imaging |
| Speaker |
Ülker Başar |
| Date |
May 28th, 2025 |
| Time |
15:30 (GMT+3) |
| Advisor |
Süha Tuna |
Abstract
Hyperspectral imaging (HSI) captures data across numerous spectral bands, offering rich information for various applications, including remote sensing, medical diagnostics, and material classification. However, the high dimensionality and vast data volumes necessitate efficient processing and compression techniques. Tensor decomposition-based methods provide a promising solution by enabling the representation of hyperspectral data in lower-dimensional components while preserving crucial information. In particular, Enhanced Multivariance product representation (EMPR) and its optimized variants have gained significant attention due to their efficiency in lossy compression and feature extraction. EMPR-based approaches decompose HSI data into hierarchical structures, reducing redundancy and computational complexity. Moreover, recent advancements integrate optimization techniques, such as the Alternating Direction Method of Multipliers (ADMM), to enhance support vector selection, further improving classification performance. This paper reviews tensor decomposition techniques in hyperspectral imaging, highlighting their applications, advantages, and challenges. The findings demonstrate that tensor based methods, particularly EMPR, play a critical role in advancing HSI processing and improving accuracy and efficiency in hyperspectral image analysis.
|
| Location |
Informatics Institute
Room 412
|
|