Projects

Biomedical image analysis plays a critical role in the diagnosis and treatment of patients with serious diseases. Advances in data availability and machine learning have improved the accuracy and efficiency of these analyses. This paper explores the use of convolutional neural networks (CNNs) for blood cell classification, a task essential for complete blood counts that identify infections based on blood cell characteristics. Manual assessment of blood cells is error-prone and resource-intensive, highlighting the need for automated solutions. We implemented a CNN using the BloodMNIST dataset, which contains over 17,000 images across eight blood cell classes. The CNN achieved superior performance with an accuracy of 0.914, significantly outperforming the baseline 1-layer Feedforward Neural Network on all evaluated metrics. Our results demonstrate that CNNs, with their ability to capture spatial hierarchies, provide a robust solution for blood cell classification, offering improved accuracy and reliability for medical diagnostics.

This paper presents a detailed study on the fine-tuning of Google’s t5 model aimed at enhancing its ability to generate high-quality, relevant questions. Leveraging the diverse and widely accepted SQuAD dataset, we trained our model specifically for question generation. Our results indicate significant improvements in multiple performance metrics, including BLEU, ROUGE and Word Mover’s Distance scores. The results were evaluated in a qualitative manner using an interactive interface. We discuss the implications of these findings for the field of model based question generation, highlighting potential applications in education. Despite the advancements, we acknowledge certain limitations in terms of generalization. Future work is suggested to further refine the model, explore alternative architectures, and integrate the enhanced model into broader AI systems.

This study explores the effectiveness of different prompting techniques for fallacy detection, a task requiring complex reasoning, using three State-of-the-Art (SotA) LLMs: Llama3, Gemma 7B, and Mistral 7B. We evaluate five prompting techniques: a handcrafted baseline, Chain-of-Thought (CoT), Thread-of-Thought (ThoT), Knowledge Generation, and Self-Consistency, against the Multi-level Annotated Fallacy Dataset (MAFALDA). The study aims to determine the impact of these techniques on LLM performance in fallacy detection, identifying the best combination of LLM and prompting technique according to the MAFALDA taxonomy. The findings highlight the individual strengths and limitations of each prompting technique and discuss future research directions.

To streamline the generation of high-quality multiple-choice questions (MCQs), two sequential pipelines were created: one with a modular design for question/answer and distractor generation, and another with a single unified model. BART was fine-tuned on the SciQ dataset, and the unified model outperformed the modular model on qualitative measures and cosine similarity of correct answers and distractors. Another pipeline was developed to extend the use of the model: given a URL and a topic, it generates an MCQ. A fine-tuned BART-large model successfully generated MCQs beyond the science domain of the SciQ dataset.

In our Dataset Bias Analysis project, we conducted a literature review and empirical study on dataset biases in image classification. Building on the work of Torralba and Efros in 2011, we reproduced their Histogram of Gradients (HOG) and Support Vector Machines (SVM) results and compared them with modern Convolutional Neural Networks (CNNs) using the caltech101, MSRC, and PASCAL VOC-2007 datasets. Our study focused on identifying selection, framing, and labeling biases, and we found that both SVM and CNN models showed significant performance drops when trained and tested on different datasets, highlighting the persistent problem of dataset bias affecting model generalization.

This project utilizes data from the Global COVID-19 Trends and Impact Survey (CTIS) and Corona data Germany to access the the impact of pandemic restrictions on mental health during COVID 19 in Germany.

In our statistical internship, we conducted multivariate analysis on climatological and hydrological low water drivers in Bavaria using simulated and measured data from three river gauges between 1990 and 2020, focusing on predicting low water events through Generalized Additive Models (GAMs) with driver-specific splines and variable-specific quantile distribution analysis, highlighting the importance of groundwater level, precipitation group, and soil water as drivers, observing distinct north-south differences, and exploring the relationships between extreme events and relevant variables.

We use the UMD Global CTIS survey, policy response information as well as Google trends data provided from Our World in Data. As part of our exploratory data analysis we built a Shiny App displaying important mental health metrics from the data.