Graduate School of Informatics

This thesis introduces a Large Language Model (LLM)-driven framework for automated augmentation policy optimization in image classification. The approach leverages LLM reasoning to generate and adapt augmentation strategies without human intervention. Two modes are explore: a before-training method that iteratively refines static LLM-generated policies, and an in-training adaptive method where policies evolve based on real-time performance feedback. By integrating semantic understanding with optimization loops, the framework tailors augmentation to dataset characteristics, enhancing model robustness and generalization compared to conventional augmentation techniques.

Date: 28.08.2025 / 14:45 Place: B-223

This thesis develops a framework for six degrees-of-freedom (6DoF) spatial audio capture and reproduction using higher-order Ambisonics and rigid spherical microphone arrays (SMAs). It introduces SPWIN (Sparse Plane-Wave Interpolation), a novel method for accurately reconstructing sound fields between microphone positions, enabling realistic, head-tracked VR experiences. The framework also presents a spherical harmonic–based method for estimating sound source directivity. Performance is validated through objective interaural coherence analysis and subjective MUSHRA tests in VR. These advancements close key gaps in spatial audio reproduction and align with emerging standards such as MPAI-CAE, supporting next-generation immersive media systems.

Date: 25.08.2025 / 13:30 Place: B-223

This study addresses critical challenges in disease progression modeling, particularly longitudinal data scarcity and class imbalance in medical imaging. The thesis proposes a novel conditional diffusion framework for synthesizing realistic disease progression sequences from cross-sectional data, utilizing ulcerative colitis endoscopic images. A diffusion model is developed with two distinct ordinal class embedding strategies that enable interpolation between discrete disease severity classes. This approach transforms static medical datasets into dynamic progression models, offering solutions for enhanced clinical training and disease understanding in medical applications.

Date: 29.08.2025 / 13:30 Place: B-223

This thesis presents a framework to enhance the quality of LLM-assisted academic peer review by addressing the lack of domain-specific evaluation criteria. While general review guidelines cover broad concerns, they often miss methodological nuances. The proposed system extracts review questions from past OpenReview evaluations and aligns them with new submissions using structured metadata such as methodology, datasets, and evaluation metrics. Experiments show that the framework improves review specificity, reduces hallucinated content, and enhances interpretability—providing more explainable and relevant automated reviews compared to baseline models.

Date: 21.08.2025 / 14:00 Place: A-212

This thesis proposes a cost-driven predictive maintenance strategy for turbofan engines using the C-MAPSS dataset. The method predicts failures within multiple horizons using LSTM models, interprets predictions via SHAP analysis, and applies maintenance decisions based on dominant sensor contributions. Various maintenance and failure cost ratios are evaluated, and strategies are compared to identify the most cost-effective approach. Results provide decision-making insights for companies with different cost priorities and operational constraints.

Date: 25.08.2025 / 15:30 Place: A-212

This study explores university students’ behavioral intention to adopt generative AI tools for learning, using an extended Technology Acceptance Model (TAM). Constructs such as trust, social influence, hedonic motivation, and task-technology fit are integrated to better understand the factors affecting adoption. Partial Least Squares Structural Equation Modeling (PLS-SEM) is applied to analyze survey data collected from students. The findings aim to provide insights into students’ perceptions and guide effective integration of generative AI in educational contexts.

Date: 27.08.2025 / 10:30 Place: A-212

This study presents a mobile authentication method based on modeling the distribution of touch behavior features. Instead of using summary statistics, the approach represents user interactions as probability distributions and compares them using KL divergence. A Siamese neural network is used to learn differences between users. The method is evaluated on a public dataset, showing improvements over baseline models in terms of error rates. Results suggest that distribution-based modeling can provide useful information for continuous authentication without requiring active user input.

Date: 21.08.2025 / 10:00 Place: A-212

This thesis investigates adaptive window sampling strategies for continuous mobile behavioral biometric authentication using accelerometer data. A dual-branch Siamese model architecture leveraging time-series and frequency-domain features is proposed to generate discriminative embeddings. Multiple triplet selection techniques—including session-balanced, session-weighted, entropy-filtered, and position-aware sampling—are explored to enhance model generalization. A window-ensembled inference strategy is employed to improve verification robustness. Experimental results demonstrate the effectiveness of targeted sampling and highlight the importance of temporally informed triplet construction in mobile biometric systems.

Date: 21.08.2025 / 11:00 Place: A-212

This study investigates how sequence variations in microexon and intronic regions influence splicing outcomes under differential SRRM4 expression. A custom CNN model was developed to predict PSI values from sequence data, followed by interpretation via DeepLIFT and motif discovery using TF-MoDISco-lite. The approach uncovered cis-regulatory patterns predictive of SRRM4-sensitive splicing, providing insights into the regulatory logic governing microexon inclusion.

Date: 01.09.2025 / 11:00 Place: A-212