|
BBL 596E/696E Seminars
CRN 23512/23514
|
|
1
|
| Title |
Multi-Agent Reinforcement Learning for Cooperative Routing, Search, and Tracking in UAV Swarms |
| Date |
April 22nd, 2026 |
| Time |
13:30 (GMT+3) |
Abstract
This seminar focuses on the use of Multi-Agent Reinforcement Learning (MARL) for cooperative decision-making in UAV swarms, with emphasis on three core tasks: routing, target search, and cooperative target tracking. Recent studies show that MARL provides an effective framework for enabling multiple unmanned aerial vehicles to make distributed decisions based on local observations while still achieving coordinated swarm-level behavior. In this context, UAV swarms are expected to operate in dynamic and uncertain environments where centralized control is often inefficient, communication is limited, and scalability is a major concern. The literature therefore increasingly investigates decentralized or partially decentralized learning strategies for swarm intelligence problems.
|
| Location |
Informatics Institute
Room 317
|
|
|
2
|
| Title |
Large Language Models for Tactical Decision Support in Air and Military Operations |
| Date |
April 22nd, 2026 |
| Time |
13:30 (GMT+3) |
Abstract
This seminar examines the emerging role of Large Language Models (LLMs) in tactical decision support for air and military operations. The focus is on three closely related functions: course of action generation, situational analysis, and command support within command-and-control environments. Recent studies indicate that LLM-based systems can accelerate the generation of operational alternatives, improve the interpretation of complex mission information, and support more flexible human–machine interaction in time-critical environments. However, their practical use remains dependent on validation, doctrinal alignment, and human oversight due to risks such as hallucination and unreliable reasoning in high-stakes settings. Overall, the seminar argues that LLMs should be treated as decision-support tools that augment, rather than replace, tactical planners and commanders.
|
| Location |
Informatics Institute
Room 317
|
|
|
3
|
| Title |
Artificial intelligence agents in cancer research and oncology paper review |
| Date |
April 22nd, 2026 |
| Time |
13:30 (GMT+3) |
Abstract
Since 2022, artificial intelligence (AI) methods have progressed far beyond their established capabilities of data classification and prediction. Large language models (LLMs) can perform logical reasoning, enabling them to plan and orchestrate complex workflows. By using this planning ability and equipped with the ability to act upon their environment, LLMs can function as agents. Here we discuss existing and emerging applications in cancer research and address real-world challenges from the perspective of academic, clinical and industrial research.
|
| Location |
Informatics Institute
Room 317
|
|
|
4
|
| Title |
Brain Tumor Segmentation from MRI Images |
| Date |
April 29th, 2026 |
| Time |
13:30 (GMT+3) |
Abstract
This study aims to investigate brain tumor segmentation from multi-modal MRI images. A U-Net–based deep learning architecture will be employed to perform voxel-wise segmentation of tumor subregions, including enhancing tumor, tumor core, and whole tumor. The proposed approach will be evaluated on a publicly available dataset using standard segmentation metrics.
|
| Location |
Informatics Institute
Room 317
|
|
|
5
|
| Title |
Automated Generation of Worst-Case Test Scenarios Using Adversarial Reinforcement Learning |
| Date |
April 29th, 2026 |
| Time |
13:30 (GMT+3) |
Abstract
Testing complex engineering components often fails to expose rare but safety-critical conditions that may occur under extreme operating regimes. To address this limitation, this study proposes an automated framework that employs adversarial reinforcement learning (RL) to generate worst-case test scenarios in a data-driven and systematic manner. The framework integrates a parameterized simulation environment constructed from historical test and operational data, enabling an RL agent to iteratively learn input and stress conditions that drive the component toward potential failure regions while respecting physical and safety constraints. The reward structure combines multiple objectives, including failure likelihood, exploration novelty, and constraint adherence. The proposed approach will be benchmarked against alternative optimization-based scenario generation techniques such as Bayesian optimization and evolutionary search. The expected outcome is an interpretable and reproducible set of critical test scenarios that can support engineers in validating component robustness and improving reliability under extreme conditions.
|
| Location |
Informatics Institute
Room 317
|
|
|
6
|
| Title |
Paired H&E-to-HER2 Virtual Staining with Conditional Diffusion Models |
| Date |
April 29th, 2026 |
| Time |
13:30 (GMT+3) |
Abstract
Compared to real immunohistochemistry (IHC), virtual staining seeks to computationally produce clinically meaningful histopathological stains, potentially lowering cost and turnaround time. In this project, paired breast cancer histopathology data will be used to investigate diffusion-based virtual staining. In particular, the publicly accessible BCI dataset will be used to train a conditional diffusion model that will synthesize HER2 IHC patches from related H&E patches. Using feature-based metrics when appropriate and coupled image quality metrics (such as PSNR and SSIM), the fidelity of the generated pictures will be compared to the ground-truth IHC. A reliable diffusion baseline on paired H&E–HER2 data with a defined evaluation process is the result of this study, and it will serve as the basis of future advances in generality and robustness.
|
| Location |
Informatics Institute
Room 317
|
|
|
7
|
| Title |
Latency-Aware Lightweight VM Scheduling for Resource-Constrained Edge Computing Environments |
| Date |
May 6th, 2026 |
| Time |
13:30 (GMT+3) |
Abstract
The proliferation of Internet of Things (IoT) devices and latency-sensitive applications has driven increasing demand for efficient computational resource management at the network edge. Edge computing environments are characterized by resource-constrained nodes with limited CPU, memory, and storage capacities, where traditional cloud-centric virtual machine (VM) scheduling approaches often fail to meet strict latency requirements. This paper proposes a latency-aware lightweight VM scheduling framework based on weighted bipartite graph matching for resource-constrained edge computing environments. The proposed approach models the VM-to-edge-node assignment as a minimum-weight bipartite matching problem, where edge weights encode a composite cost function incorporating estimated task completion latency, resource utilization efficiency, and network proximity. We employ a modified Hungarian algorithm with O(n³) time complexity to compute optimal assignments, making it suitable for real-time scheduling decisions at the edge. A comparative analysis against existing scheduling strategies including Round Robin, Min-Min, First-Fit Decreasing, and reinforcement learning-based approaches is presented across key metrics such as computational complexity, latency sensitivity, and scalability. The analysis demonstrates that the proposed bipartite matching approach offers a favorable balance between scheduling optimality and computational overhead for small-to-medium scale edge deployments.
|
| Location |
Informatics Institute
Room 317
|
|
|
8
|
| Title |
A Comparative Study of Fuzzy Logic and Round
Robin Load Balancing in Software-Defined
Networks |
| Date |
May 6th, 2026 |
| Time |
13:30 (GMT+3) |
Abstract
Software-Defined Networking (SDN) enables centralized network control, making load balancing a critical
challenge. This paper provides a comparison-based analysis for the traditional Round Robin load balancing algorithm with a fuzzy logic-based intelligent approach in SDN. The proposed system is developed based on the Ryu controller and evaluated in a Mininet emulation environment with three backend servers. The simulation results are used to validate both algorithms on larger-scale experiments, where we verify that the throughput
and the response time are similar between both approaches (~351 requests/sec) under static conditions, while adapting in the Fuzzy Logic strategy provides a smart load distribution with servers with higher latency can be relieved from load on loaded servers by which this approach can handle better at adaptation to dynamic conditions.
|
| Location |
Informatics Institute
Room 317
|
|
|
9
|
| Title |
Instance Segmentation and Multi-Object Tracking in CAMAD Microscopy
Datasets using YOLO-SAM Hybrid Pipeline |
| Date |
May 6th, 2026 |
| Time |
13:30 (GMT+3) |
Abstract
Automated cell tracking in time-lapse microscopy remains a significant challenge due to high cell density, non-rigid deformations, and low signal-to-noise ratios in phase-contrast imaging. Traditional semantic segmentation (e.g., standard U-Net) often fails to separate ”clumped” or touching cells, leading to identity switches and tracking failures. This project addresses these challenges by implementing an instance-first tracking pipeline on
the CAMAD dataset.
|
| Location |
Informatics Institute
Room 317
|
|
|
10
|
| Title |
Edge Computing (It is not fully decided yet) |
| Date |
May 13th, 2026 |
| Time |
13:30 (GMT+3) |
Abstract
The amount of data that needs to be processed by mobile users is increasing daily, requiring high-speed communication to reduce latency and greater computing capabilities. Edge computing ffers an effective solution. Edge computing is designed to process users’ tasks faster and utilize greater computation capabilities. Compared to traditional cloud computing, computation resources are located closer to users. With
the development of next-generation mobile systems and the
emergence of new application areas, the need for edge computing
solutions is increasing. Solutions and models proposed for the task
offloading problem play a crucial role in realizing these systems.
In this paper, multiple edge servers and the main cloud coexist,
and tasks can be offloaded to one of them or computed locally
on the mobile user. Round-robin task scheduling is performed
on edge servers, and mobile users generate tasks using a Poisson
distribution and follow a uniform distribution. The action with the lowest sojourn time is selected for the offload decision. Compared to the offload process structured in most optimization problems in the literature, this paper constructs a continuous and stochastic model. Energy consumption calculations are used to examine the system’s efficiency.
NOTE: Paper topic is not fully decided yet so topic and abstract may change over time.
|
| Location |
Informatics Institute
Room 317
|
|
|
11
|
| Title |
Diagnose Like A REAL Pathologist: An Uncertainty-Focused Approach for Trustworthy Multi-Resolution Multiple Instance Learning
This article proposes the Uncertainty-Focused Calibrated Multiple Instance Learning (UFC-MIL) model, which mimics the behavior of pathologists who focus on uncertain areas and zoom in at different resolutions, and is developed to provide reliable, well-calibrated results in histopathological diagnoses. |
| Date |
May 13th, 2026 |
| Time |
13:30 (GMT+3) |
Abstract
With the increasing demand for histopathological specimen examination and diagnostic reporting, Multiple Instance Learning (MIL) has received heightened research focus as a viable solution for AI-centric diagnostic aid. Recently, to improve its performance and make it work more like a pathologist, several MIL approaches based on the use of multiple-resolution images have been proposed, delivering often higher performance than those that use singleresolution images. Despite impressive recent developments of multiple-resolution MIL, previous approaches only focus on improving performance, thereby lacking research on well-calibrated MIL that clinical experts can rely on for trustworthy diagnostic results. In this study, we propose Uncertainty-Focused Calibrated MIL (UFC-MIL), which more closely mimics the pathologists’ examination behaviors while providing calibrated diagnostic predictions, using multiple images with different resolutions. UFC-MIL includes a novel patch-wise loss that learns the latent patterns of instances and expresses their uncertainty for classification. Also, the attention-based architecture with a neighbor patch aggregation module collects features for the classifier. In addition, aggregated predictions are calibrated through patch-level uncertainty without requiring multiple iterative inferences, which is a key practical advantage. Against challenging public datasets, UFC-MIL shows superior performance in model calibration while achieving classification accuracy comparable to that of state-of-the-art methods.
|
| Location |
Informatics Institute
Room 317
|
|
|
12
|
| Title |
Computational Interpretation of Gain- and Loss-of-Function Mutations: A Systematic Survey |
| Date |
May 13th, 2026 |
| Time |
13:30 (GMT+3) |
Abstract
In the era of precision medicine, distinguishing between Gain-of-Function (GoF) and Loss-of-Function (LoF) mutations is essential, as these opposing mechanisms often require fundamentally different therapeutic strategies. Although computational tools can reliably predict variant pathogenicity, determining the specific functional direction of mutations remains a significant challenge. This systematic review addresses this critical gap by highlighting the conceptual transition from binary pathogenicity assessment toward mechanism-specific classification.
The computational landscape between 2015 and 2025 is examined, and a comprehensive taxonomy is introduced, categorizing existing approaches into biological infrastructure–based methods, feature-based learning, deep learning models, and phenotype-driven frameworks. In addition to this conceptual taxonomy, a multidimensional comparative analysis is presented to evaluate trade-offs among feature dimensionality, dataset scale, and predictive performance.
Quantitative benchmarking indicates a clear divergence in computational strategies: early approaches relied on high-dimensional, brute-force feature representations, whereas recent state-of-the-art models demonstrate that optimized, high-signal feature sets—particularly those derived from AlphaFold structural predictions—can achieve superior performance with reduced model complexity. The analysis further highlights the persistent data scarcity bottleneck, characterized by severe class imbalance in which experimentally validated GoF variants are substantially underrepresented compared to LoF variants. This imbalance limits the generalization capacity of highly data-dependent deep learning models.
Additionally, a label validity crisis within current biological infrastructure is identified, where historical misannotations (e.g., TP53 variants) and hidden class imbalances compromise benchmark reliability. The review concludes by outlining a future research roadmap emphasizing hybrid computational architectures that integrate genomic representation learning with clinically informed priors, aiming to bridge the gap between computational prediction and clinical applicability.
Keywords—Computational Variant Interpretation, Gain of-Function/Loss-of-Function, Structural Genomics, Machine Learning, Precision Medicine, Mechanism Prediction.
|
| Location |
Informatics Institute
Room 317
|
|
|
13
|
| Title |
Real-Time Driver Distraction Detection Using Deep Learning for Enhanced Road Safety |
| Date |
May 13th, 2026 |
| Time |
13:30 (GMT+3) |
Abstract
Driver distraction is a leading cause of traffic accidents, making in-cabin monitoring essential for modern intelligent vehicles. This seminar presents an automated distraction detection system using Convolutional Neural Networks (CNNs). By analyzing dashboard camera images, the proposed model classifies driver behaviors into distinct categories, such as safe driving, texting, or operating the radio. The methodology leverages transfer learning with lightweight architectures like MobileNet to achieve high real-time accuracy suitable for edge deployment. Covering the end-to-end pipeline from data augmentation and preprocessing to model evaluation using public datasets, this framework offers a practical and scalable computer vision solution to enhance Advanced Driver-Assistance Systems (ADAS) and overall road safety.
|
| Location |
Informatics Institute
Room 317
|
|
|
14
|
| Title |
Entropy-Based Resource Control for Digital Twin Synchronization |
| Date |
May 20th, 2026 |
| Time |
13:30 (GMT+3) |
Abstract
Digital twins are powerful tools used to model and predict the behavior of physical systems. However, their accuracy often decreases over time due to changing physical conditions, inconsistent measurement sources, and limited system resources. This study analyzes the information gap between measurements from the physical system and its digital twin, focusing on how this gap evolves over time and affects resource usage.
The framework evaluates the contribution of each system component (such as sensors or network devices) to overall information accuracy and efficiency, considering resource consumption such as CPU, bandwidth, and memory. It also introduces a fidelity metric to measure the accuracy between the physical and digital systems.
In addition to global information mismatch, the model considers structural differences within the system state to detect hidden inconsistencies. This unified framework enables evaluation of temporal accuracy, resource allocation, and model update needs in digital twins, and proposes an information-centric approach for improving future model adaptation and efficiency.
|
| Location |
Informatics Institute
Room 317
|
|
|
15
|
| Title |
Regime-Dependent Fuzzy Inference System for Quantifying the Impact of Economic Advisors on S&P 500 Performance: Fuzzy Advisor Impact Index - FAII |
| Date |
May 20th, 2026 |
| Time |
13:30 (GMT+3) |
Abstract
Financial sentiment analysis traditionally focuses on retail investor noise, often neglecting the structural market impact of macroeconomic policymakers. This paper introduces the Fuzzy Advisor Impact Index (FAII), a novel trading system that quantifies the linguistic influence of 39 key economic figures—including Federal Reserve Chairs and Treasury Secretaries—using news data scraped from The Guardian (2009–2026).
The model employs a Dynamic Active-Roster Weighting mechanism to filter inactive advisors and processes signals through a Type-1 Fuzzy Inference System. Crucially, the system integrates an Inflation Regime Filter based on Crude Oil trends to address the ""Deflation Trap,"" dynamically adjusting the model's sensitivity to bearish signals based on the inflationary environment.
Out-of-sample testing (2013–2026) demonstrates that the regime-aware FAII strategy delivers superior performance, achieving a Total Return of 426.9% compared to the S&P 500 Buy-and-Hold benchmark’s 368.8%. Furthermore, the strategy significantly improves risk-adjusted returns, attaining a Sharpe Ratio of 1.04 versus the benchmark’s 0.82. These results suggest that policy sentiment is a robust source of alpha when conditioned on macroeconomic regimes.
|
| Location |
Informatics Institute
Room 317
|
|
|
16
|
| Title |
Rule-Based Generation of Synthetic Chromosome Images for Mechine Learning Applications |
| Date |
May 20th, 2026 |
| Time |
13:30 (GMT+3) |
Abstract
Background: In cytogenetics, deep learning-based projects require large annotated chromosome datasets, which constitutes a major limitation due to the time-consuming and costly nature of annotation. To the best of our knowledge, this study is the first to employ a rule-based chromosome image generation approach with controllable morphological structures to address dataset limitations in cytogenetics for machine learning applications. The proposed approach aims to provide chromosome images as a reliable data source and reduce dependency on natural data. Furthermore, it can be extended to generate chromosomes with abnormalities, supporting anomaly detection tasks.
Methods: Rule-based chromosome generation framework was developed based on idiogram-derived structural information and mathematical modeling, enabling the creation of annotated synthetic chromosome images. Chromosome-specific morphological and textural parameters were autonomously optimized using Bayesian optimization, where the classification accuracy of a CNN trained on real chromosome images was used as the objective measure to assess the realism of the generated synthetic chromosomes.
Results: Cross-domain evaluation showed that classification accuracy reached 98.72% when models trained on real chromosome images were tested on synthetic data, indicating high structural consistency. When models trained on synthetic chromosome images were evaluated on real data, classification accuracy reached 93.46%, with an overall F1 score of 92.07% across 24 chromosome classes. These findings indicate that the generated synthetic images effectively capture key morphological features and class-specific characteristics of real chromosomes.
Conclusion: Rule-based synthetic chromosome generation is an effective data generation strategy for cytogenetic machine learning applications.
|
| Location |
Informatics Institute
Room 317
|
|