From: 2024-10-31 11:15 to 17:00
Place: See information for each presentation
Contact: birger [dot] swahn [at] cs [dot] lth [dot] se

Five MSc and BSc theses to be presented on Thursday October 31, 2024

Thursday October 31 is a day for coordinated master and bachelor thesis presentations in Computer Science at Lund University, Faculty of Engineering. Five MSc and BSc theses will be presented.

You will find information about how to follow along under each presentation. The presentations will be in two different rooms: E:2116 and E:2405 (Glasburen). See information for each presentation. A preliminary schedule follows.

Note to potential opponents: Register as an opponent to the presentation of your choice by sending an email to the examiner for that presentation (firstname.lastname@cs.lth.se). Do not forget to specify the presentation you register for! Note that the number of opponents may be limited (often to two), so you might be forced to choose another presentation if you register too late. Registrations are individual, just as the oppositions are! More instructions are found on this page.

11:15-12:00 in E:2405 (Glasburen)

Presenter: Johan Ravnborg
Title: Performance Optimisation of Collision Detection Algorithm in Particle Simulation
Examiner: Michael Doggett
Supervisor: Rikard Olajos (LTH)

This thesis evaluates the effect of different optimisation steps on a collision detection algorithm, applied to a dynamic particle simulation. Starting from a basic implementation, progressive optimisations were added to the original code, as well as large reworks of the structure, testing different basic Bounding Volume Hierarchies. A 2D physics engine was created with monosized particles delimited by Oriented Bounding Boxes. Gravity was applied towards the overall mass center to facilitate collisions. To understand the performance improvement of each step, measurements were made for different optimisation steps of the collision detection algorithm over varying numbers of particles. We found that bounding circle pre-testing for equal-sided Oriented Bounding Boxes saved significant amounts of demanding overlap testing. We also found that Top Down Bounding Volume Hierarchies using cheap build processes such as median split were highly efficient in dynamic simulations with multiple checked entities.

Link to popular science summary: To be uploaded

13:15-14:00 in E:2116

Presenter: Charles van Amersfoort
Title: Simplifying Embedded Systems with a Rust Manifest for Multi-Language Services
Examiner: Emma Söderberg
Supervisors: Görel Hedin (LTH), Olof Winge (Schneider Electric)

Event-driven embedded systems face challenges in managing cross-language communication when multiple services are developed using different programming languages. This thesis introduces a Rust-based manifest solution to standardize message definitions and interfaces across services, improving maintainability, type safety, and developer experience. By utilizing Rust's strong type system and macros, the proposed declarative approach simplifies integration and enhances efficiency in multi-language environments, particularly in embedded systems. This solution is evaluated through the case study of Schneider Electric's Fire Detection Panel, where it demonstrates a significant improvement in cross-language interoperability, system scalability, and development speed, providing a robust alternative to existing methods like AsyncAPI.

Link to popular science summary: To be uploaded

14:15-15:00 in E:2116

Presenter: Kadriye Yildiz
Title: Assessing AI’s Problem Solving Capabilities
Examiner: Görel Hedin
Supervisors: Emma Söderberg (LTH), Andreas Bexell (LTH)

As AI continues to evolve, there is growing interest in its potential to either replace or assist software engineers. This thesis examines the problem-solving abilities of GPT-4o, using the Kattis platform to automatically solve 4,237 coding problems. Two temperature settings (0.1 and 0.5) were tested for consistency, and problem characteristics like description length and difficulty were analyzed. GPT-4o achieved an average acceptance rate of 23%, mainly on easy problems, with more consistent performance at temperature 0.1. Accepted solutions had shorter descriptions (330 vs. 420 words) and lower difficulty levels (3.11 vs. 5.27). While GPT-4o performed similarly to Kattis users on easy problems, it was significantly outperformed on medium and hard tasks. Despite producing 65.81% shorter code on average, its overall success rate was lower. These findings suggest that GPT-4o can assist in automating simple programming tasks but struggles with more complex challenges.

Link to popular science summary: To be uploaded

15:15-16:00 in E:2405 (Glasburen) and in Teams Meeting ID: 318 252 686 561 Passcode: u2dvVX (see link below) N.B Bachelor thesis

Presenter: Robin Baki Davidsson
Title: Quantization of Large Language Models and the Impact on Output Performance
Examiner: Noric Couderc
Supervisor: Pierre Nugues (LTH)

This thesis investigates the performance of eight large language models (LLMs) at various quantization levels, focusing on tasks such as MMLU-Pro for knowledge and reasoning, CRUXEval for code comprehension, and MuSR for reading comprehension and coherence. The results demonstrate a consistent trend where higher bit precision (e.g., 8-bit Q8_0) yields improved performance, albeit with diminishing returns. Aggressive quantization (e.g., 2-bit Q2\_K) usually retains acceptable accuracy, though some models may occasionally lose substantial performance and coherence. The findings indicate that while lower bit precision generally reduces performance, the impact varies across models and tasks. Larger models show greater resilience to aggressive quantization but can still experience significant performance drops at lower precision levels. Mid-sized models in the 7-9 billion parameter range strike an optimal balance between capability and efficiency. These results provide valuable insights into the trade-offs between model size, quantization, and task performance in real-world applications.

Link to popular science summary: To be uploaded

Link to Teams presentation: https://teams.microsoft.com/l/meetup-join/19%3ameeting_MzE5NzNlZDUtZjU0ZC00ODgxLWE2YmEtNDM3MjMwZjIxYmRi%40thread.v2/0?context=%7b%22Tid%22%3a%227aa68094-6104-41a6-b443-d4b52451f617%22%2c%22Oid%22%3a%22236ccee7-a8cc-4808-b395-d4f4d11d1f6f%22%7d

16:15-17:00 in E:2405 (Glasburen)

Presenter: Edvin Alicajic
Title: AI text models for biodiversity data analysis to guide sustainable development
Examiner: Flavius Gruian
Supervisor: Pierre Nugues (LTH)

Our perception of nature is believed to be one of the reasons why biodiversity has been negatively affected by us. By using tools such as natural language processing and AI-text models, we might be able to gain more knowledge about this subject and guide sustainable development forward. This thesis uses four different approaches to analyze newspapers between 1920-2020: Frequency analysis, Co-occurrences analysis, sentiment analysis, and topic modeling. We found that all four approaches managed to tell us how our perception has changed. Our language use has simplified, and the general sentiment around birds has become more polarized. The focus of associated birds has changed from individual traits towards a more holistic view. We also saw that nature and culture still have a connection, but one that has changed over time. In conclusion, AI text models can be used efficiently to collect data to answer questions related to sustainable development.

Link to popular science summary: To be added