Index

Dynamic Query Routing: Adaptive RAG Systems Leveraging Hallucination Risk and Specialization Affordance

Large Language Models responses are often generic and may contain false information leading to hallucinations. It takes a lot of resources to train LLM with updated data. As such, Individual organizations can use a Retrieval Augmented Generation (RAG) system with LLM to avoid those drawbacks and make it their own. It uses a private knowledge base to pick up necessary information relevant to the user’s query and feed it to LLM to generate detailed and context rich responses. However, RAG system may give rise to conflict between LLM’s prior learned memory and in-context memory due to new resources.

RAG can be optimized based on how the external knowledge base is utilized. Focus on key extraction from query, structured knowledge base, iteration/ recursion in searching relevant information and re-ranking mechanisms based on knowledge density/diversity are some of the strategies that can be used. Among them, there is a strategy that analyzes the user’s query if it is a simple one that requires smaller context or can be answered without external knowledge base, or complex one that requires more context to get accurate answers. This leads to dynamically adapting the retrieval process according to the user’s need and adjusting the context window. It can also refine its performance based on user interaction and feedbacks by prioritizing resources that have previously provided accurate and relevant answers.

This thesis will focus on development of similar approach and work on comparing them with previous similar works and with conventional RAG methods and improve performance where possible. Recent updates of LLMs such as GPT series, BERT or T5 models, which provide strong foundational capabilities for text generation and can be used for fine-tuning key extraction tasks, may be used depending on available computational resources. The research will use a digital copy of one of the course books used in Masters in AI program, as the primary source for external knowledge base, and use query datasets/evaluation datasets (a mix of straightforward queries and conceptual questions) based on the book. Recent similar works include the following:

Jeong, S., Baek, J., & Cho, S. (2024). Adaptive-RAG: Learning to Adapt Retrieval-Augmented Large Language Models through Question Complexity. https://arxiv.org/pdf/2403.14403v2
Wang, X., Sen, P., & Li, R. (2024). Adaptive Retrieval-Augmented Generation for Conversational Systems. https://arxiv.org/pdf/2407.21712

Improving AFFGANwriting by Exploring Deep Learning Models for Style Encoders and Image Generation of Sentence-Level Handwriting

Handwriting generation is a fundamental task in computer vision and natural language processing, with applications in personalized content generation and so on. The AFFGANwriting model presents a generative framework for synthesizing word-level handwritten images by fusing multistyle features using a GAN-based approach with a VGG-style encoder. However, its scope is limited in two ways:
• It only generates individual word images
• It used a fixed VGG backbone which may not capture style semantics as effectively as more modern alternatives such as CNN and transformer models (e.g. EfficientNet, ResNet, DINO).

With an increasing demand for personalized handwriting synthesis across longer text spans, there’s a clear motivation to explore if advanced backbone models can improve the feature extraction of the style. In addition, there is need to extend the generative capacity from words to full sentences and to interact ideally in a user-friendly interactive system.

Research questions

• Can more recent feature extractors like CNN and transformers (EfficientNet, ResNet, DINO) outperform VGG in capturing style-relevant features for handwriting generation?
• What are the architectural or training modifications required to extend AFFGANwriting from word-level to sentence-level image synthesis?
• How can the model be integrated into an intuitive web application that allows users to select a writing style and input arbitrary text for sentence-level generation?

Goal

To enhance AFFGANwriting’s quality and flexibility in handwriting image generation by:
• Upgrading the style encoder
• Enabling sentence-level synthesis
• Deploying the system as a web app for user interaction

Plug-and-Play Diffusion Models for Magnetic Resonance Imaging

Development of an AI-Based Algorithm for the Correction of Moiré Artifacts in Digital Radiography

Synthetic Non-Contrast CT Angiography Image Generation using Deep Learning Methods

RPA-Bots zur Prozessautomatisierung im Workflow Management der DATEV eG

Advanced Machine Learning Models for Leakage Detection and Localization in Water Distribution Networks Using Real-System Data

Reinforcement Learning for Centralized Fault Coordination in Power Systems

Latent Space Modeling for Event Detection in Power Grid Data

This project explores how latent representations learned from raw grid waveforms can reveal underlying structure and enable early detection of abnormal events. By modeling high-frequency voltage and current signals, we aim to distinguish critical disturbances from normal behavior with minimal delay.

Report Generation in pathology using WSIs

This project focuses on developing methods for processing large-scale digital pathology datasets and extracting meaningful features from whole slide images to support automated report generation. Emphasis is placed on efficient handling of gigapixel image data and preparing it for use in vision-language models for clinical applications.