Graduate School of Informatics

This research addresses the multifaceted nature of Technical Debt (TD) in Machine Learning (ML) projects, distinct from traditional software projects due to their probabilistic nature and data dependency. The study systematically examines how TD manifests across various dimensions in ML projects, identifying root causes, impacts, and band-aid solutions contributing to its persistence. ML-specific TD was categorized through thematic analysis of interviews with industry professionals. The findings were reviewed by academic experts in multiple iterations. This study fills a gap in the literature and offers practical insights for managing TD in ML contexts, as well as a TD-oriented structure for its assessment.

Date: 06.09.2024 / 09:30 Place: A-212

This thesis investigates cross-linguistic differences in realizing discourse relations, centered on the TED-MDB corpus. By developing a framework for aligning discourse relation annotations in parallel corpora, the study explores variations in discourse relation realization, semantic shifts, and inter-sentential encoding patterns across languages. Key findings highlight the importance of discourse relation linking, revealing differences in the translation of discourse connectives. Also, this study develops method for bilingual lexicon induction from aligned data, supporting pragmatic studies and natural language processing systems. Future work includes adapting discourse relation-aligned data to Linked Language Open Data (LLOD) standards for better accessibility and interoperability.

Date: 04.09.2024 / 10:00 Place: A-212

Humans have long been curious about other minds; a fascination rooted in ancient philosophy that shapes key debates in modern cognitive science. As artificial intelligences become more widespread, the human tendency to attribute mental states—known as mind perception—to non-human entities has found relevance in human-robot interaction. This thesis explores the dynamics of mind perception in this context, focusing on determinants related to both the perceiver and the perceived entity. In multiple-step experiments involving 160 participants from four generations, we examined how agent type (human vs. robot), action type (communicative vs. noncommunicative), individual traits, and generational differences influence mental capacity attributions. We measured mind perception along Agency (ability to do) and Experience (ability to feel) dimensions. Our methodology combined implicit and explicit tasks in a real-time, naturalistic lab setting with live actors, enhancing ecological validity while maintaining experimental control. We collected both behavioral measurements and self-report answers, addressing the recent discussions in the field. Findings revealed that the human was consistently attributed higher mental capacities than the robot. Action type effects were varied and context-dependent. Young participants were more likely to attribute mental states to the robot, while individual traits showed weak influences. Explicit measures aligned with implicit ones in showing higher mind attribution to the human, while implicit measures revealed subtler effects, particularly for action type and agency. Response times and mouse trajectories captured nuances and interactions that were not apparent in explicit ratings. This thesis highlights the significance of considering both perceiver- and perceived-related factors and using implicit and explicit assessment methods to reveal layered interactions among determinants by uncovering distinctions between these two measurement types.

Date: 03.09.2024 / 11:00 Place: A-212

Atmospheric turbulence, caused by factors such as temperature, wind speed, and humidity, leads to random fluctuations in the atmosphere's refractive index. This phenomenon degrades the image quality of long-range observation systems through geometric distortions and spatial-temporal varying blur. Turbulence can affect various imaging spectra, including visible and thermal bands. This thesis addresses the challenge of atmospheric turbulence in thermal imagery and its impact on object detection models. To tackle this challenge, we propose a data augmentation method that enhances the performance of object detectors by utilizing turbulent images with varying severity levels as training data. We generate training samples using a geometric turbulence simulator and use Geometric, Zernike-based, and P2S-based simulators to create the turbulent test sets, confirming the effectiveness of our augmentation method across different types of simulated turbulence. Our results demonstrate that this data augmentation approach significantly improves performance for both turbulent and non-turbulent thermal test images.

Date: 03.09.2024 / 13:30 Place: B-116

The quantization of neural networks is essential to meet real-time requirements. Zero-shot quantization is a key approach when training data is unavailable. To the best of our knowledge, zero-shot quantization in the infrared domain has not been explored before. This thesis examines the performance of batch normalization statistics-based zero-shot quantization on models trained with infrared imagery. We fine-tuned models pretrained on RGB images using infrared images and carefully investigated the data generation process to achieve optimal results for YOLOv8 and RetinaNet. Our results demonstrate that zero-shot quantization is more effective in the infrared domain.

Date: 03.09.2024 / 11:00 Place: B-116

This thesis explores efficient pretraining methods for Vision Transformers by integrating progressive layer freezing with local masked image modeling. The study assesses the computational demands and extended training periods typical of self-supervised learning methods for ViTs. Key innovations include implementing the FreezeOut method within the LocalMIM architecture to significantly enhance training efficiency. Experimental results show a reduction in training time by about 12.5% while maintaining competitive accuracy, demonstrating the effectiveness of strategic layer freezing combined with tailored learning rate scheduling. This approach promotes more accessible self-supervised learning on constrained computational resources.

Date: 03.09.2024 / 09:30 Place: B-116

This study is about narrative comprehension. Narrative comprehension involves how the characters and events described in narratives are represented in readers’ mind, what kinds of inferences are activated while reading narratives, and what text factors affect comprehension of narratives. In particular, our research will concentrate on the comprehension of spatial cues and protagonist’s perspective in narratives. Several studies in the literature suggest that readers are sensitive to the spatial cues and can, under some conditions, adopt the perspective of the protagonist in narratives. For this thesis, we will investigate this issue further, by testing some new research questions.

Date: 06.09.2024 Place: II-06

This thesis introduces advanced methodologies for Land Use Mix (LUM) analysis in urban planning, GIS research, and disaster risk assessment. It addresses limitations in traditional approaches by leveraging point-based geospatial data and develops an open-source Python package, landusemix, for scalable and adaptable LUM calculation. The research extends LUM analysis to evaluate temporal variations in urban vulnerability, particularly concerning earthquake risk, offering insights for time-sensitive urban planning. This work enhances sustainable, resilient, and livable cities through innovative tools and approaches in urban studies.

Date: 03.09.2024 / 17:00 Place: B-116

Perimeter security is no longer barrier to access networks and critical data, making traditional security measures outdated. A robust security strategy is crucial to prevent and detect Active Directory (AD) attacks, which can be difficult to detect due to their blend in with normal network traffic. One such attack is the Kerberoasting attack, which exploits weaknesses in the Kerberos authentication protocol. To detect these attacks, supervised machine learning algorithms are being proposed. And also publicly available dataset to measure the efficiency of these algorithms for Kerberoasting attacks was created and shared.

Date: 05.09.2024 / 10:00 Place: II-06

This work investigates the meaning of the perfective/evidential suffix -mIş, focusing on its perfect interpretation. It has been argued that there are two distinct syntactic structures for simple verbal sentences [verb+past] and complex verbal sentences [verb+part+cop+past] (Kornfilt, 1996; Kelepir, 2001). Demirok and Sağ (2023) offer a compositional account for these two structures, taking the temporal relations as the basis. Building on that, we propose an Aktionsart-oriented analysis of the verb-participle relation. We offer a continuation-based compositional account within quantificational event semantics (Champollion, 2015) to reconcile the syntactic account of Kelepir (2001) and observations on the perfect meaning of -mIş.

Date: 06.09.2024 / 15:30 Place: A-212