lunduniversity.lu.se

Computer Science

Faculty of Engineering, LTH

Events

CS MSc Thesis Presentation Day June 1 2023

Föreläsning

From: 2023-06-01 09:15 to 17:00
Place: See information for each presentation
Contact: birger [dot] swahn [at] cs [dot] lth [dot] se
Save event to your calendar


Eight MSc theses to be presented on Thursday June 1, 2023

Thursday June 1 is a day for coordinated master thesis presentations in Computer Science at Lund University, Faculty of Engineering. Eight MSc theses will be presented.

You will find information about how to follow along under each presentation. There will be presentations in two different rooms: E:2116 and E:4130 (Lucas). A preliminary schedule follows.

Please note that there will also be five thesis presentations on Friday June 2, schedule at: https://cs.lth.se/kalendarium/?evenemang=extra-cs-msc-thesis-presentation-day-june-2-2023

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.


 

E:2116

 

09:15-10:00 in E:2116

Presenter: Elmer Dellson
Title: Dynamic Diffuse Global Illumination Using Probes and Surfels
Examiner: Per Andersson
Supervisors: Michael Doggett (LTH), Calle Lejdfors (AMD)

One of the major challenges of real-time computer graphics is that of global illumination. This thesis explores an approach to dynamic diffuse GI based on light probes and surfels combined with hardware accelerated ray tracing, to achieve convincing second-order illumination in 3D scenes with real-time frame rates. Using ray tracing, the system places surfels that feed light information to the probes, that then propagate it across the scene. The system is dynamic and does not require any pre-baking. The results show that this method can be very fast, and yield good second-order lighting for the simple scenes sampled. We also saw that the caching of surfels saves greatly on the computational cost of the system compared to ray tracing to replace them every frame. The system predictably slows down when the number of surfels is increased, but at a manageable rate.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/230601_09Dellson.pdf

Link to Zoom presentation: https://lu-se.zoom.us/j/64768038935?pwd=R0lUQkw4ZWNnNjNkczVwcWgxUWdRZz09


10:15-11:00 in E:2116

Presenters: Uros Tripunovic, Andrej Simeunovic
Title: Managing Micro Frontends Across Multiple Tech Stacks - Sharing, Finding & Publishing
Examiner: Per Andersson
Supervisor: Lars Bendix (LTH)

IKEA, a large enterprise with various tech stacks across multiple teams, seeks to address code duplication and communication overhead by implementing Micro Frontends (MFEs) to enable code reusability and autonomous development. However, aligning technologies among teams presents a challenge. Although the teams at IKEA currently face no issues in finding and publishing MFEs due to being in the early stages, it's crucial to address this issue in advance as this will minimize duplication and enhance code reusability. Two methods were discovered for sharing MFEs at IKEA without requiring changes to the existing tech stack. These methods involve either wrapping framework components as Web Components or exposing a framework's render method within a function and injecting it into the frontend application. Each integration method varies in terms of the effort required. In addition, IKEA already has a platform for finding and publishing MFEs. To improve the ease of finding MFEs, we recommended enhancing the search functionality by displaying the teams involved in creating each MFE.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/230601_10SimeunovicTripunovic.pdf


11:15-12:00 in E:2116

Presenters: Johannes Aronsson, David Björk
Title: Extending the ExtendJ Java Compiler
Examiner: Görel Hedin
Supervisor: Idriss Riouak (LTH)

ExtendJ is a Java compiler supporting Java versions from 4 to 8, and it is built using the JastAdd metacompiler. ExtendJ is designed to enable modular extensions. This thesis aims to examine ExtendJ’s extendibility and performance by attempting to add support for Java versions 9, 10 and 11. Many features were introduced in these versions, including local type inference with the var identifier. The implemented features were then evaluated by compiling real-world projects to validate the implementation and measure compilation time as well as memory usage. Finding and compiling relevant projects proved difficult, and almost only projects using Java 8 features and earlier where compiled. The performance of ExtendJ versions 8 to 11 was compared with the corresponding OpenJDK compilers by measuring compilation time and memory consumption. The compilation time of ExtendJ was found to be within a factor of 3, while the memory consumption was within a factor of 6. We also found that ExtendJ is modularly extensible to a high degree.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/230601_11AronssonBjörk.pdf


13:15-14:00 in E:2116

Presenters: Björn Magnusson, Erik Amirell Eklöf
Title: Improvements to planar convex hull algorithms through theoretical and empirical analysis
Examiner: Michael Doggett
Supervisor: Jonas Skeppstedt (LTH)

The task of computing the convex hull given a set of points in the plane is a well known problem for which there are many algorithms with various performance related trade offs. In this thesis we aim to gain a better understanding of how existing algorithms perform in different circumstances. We have concluded that for most densely distributed inputs, Quickhull is the fastest algorithm. We have also empirically evaluated different variants of Quickhull, as well as special cases where Quickhull performs poorly compared to other algorithms. Furthermore we present a novel algorithm that combines the best aspects of Chan's algorithm and the divide and conquer algorithm by Bentley and Shamos.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/230601_13MagnussonAmirellEklöf.pdf


15:15-16:00 in E:2116

Presenter: Steven Chen
Title: User-Centric Study and Enhancement of Python Static Code Analysers
Examiner: Martin Höst​​​​​​​
Supervisors: Emma Söderberg (LTH), Alan Mccabe (LTH)

Despite the growing integration of code analysis tools in developer workflows, usability challenges persist in many aspects. Past researches, primarily focused on investigating static languages and professional developers, has overlooked the needs of novice Python developers. This thesis investigates the experiences of beginner Python programmers with static code analysis tools. We aim to understand how these beginners interact with and perceive these tools, with a focus on identifying usability pain points. The insights derived from this study will be used to enhance the Pylint extension on Visual Studio Code, incorporating additional quick-fixes intended to improve user experience. This research contributes to the field by providing a user-centric perspective on the design and functionality of Python code analysis tools for novice developers.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/230601_15Chen.pdf


 

E:4130 (Lucas)

 

10:15-11:00 in E:4130 LUCAS

Presenter: Anton Fristedt
Title: Optimizing Reinforcement Learning Algorithms Using Design Of Experiments​​​​​​​
Examiner: Elin Anna Topp​​​​​​​
Supervisor: Linda Hartman (LTH)

In this project, we conducted design of experiments on the Proximal Policy Optimization (PPO) reinforcement learning algorithm, with the aim of optimizing its performance for robotic learning tasks. We considered different hyperparameters of PPO as factors and used a factorial design to explore their effects on the algorithm's performance. We specifically focused on the CartPole-V1 task in the OpenAI Gym classic control environment and measured the time it took for the learning to reach 95\% success as the primary performance metric. Our results show that some hyperparameters have a significant impact on PPO's performance, while others have little effect. Overall, our project demonstrates the usefulness of design of experiments for optimizing RL algorithms for robotic learning tasks and provides insights into the hyperparameter tuning process.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/230601_10Fristedt.pdf


14:15-15:00 in E:4130 LUCAS (Also online in Zoom)

Presenters: Josefin Gustafsson, Pontus Rosqvist​​​​​​​
Title: Skill-Based Autonomous Door Opening​​​​​​​
Examiner: Jacek Malec​​​​​​​
Supervisor: Volker Krueger (LTH)

This project implements a door opening skill for an autonomous robot. It covers how to detect the door opening mechanism, opening the door and then going through it. This is achieved using ROS and SkiROS on a mobile robot called Heron.

Zoom link to presentation: https://lu-se.zoom.us/j/61482330273

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/230601_14GustafssonRosqvist.pdf


16:15-17:00 in E:4130 LUCAS (Also online in Zoom)

Presenters: Matilda Froste, Mosa Hosseini
Title: Multi-Label Toxic Comment Classification Using Machine Learning: An In-Depth Study​​​​​​​
Examiner: Jacek Malec​​​​​​​
Supervisors: Pierre Nugues (LTH), Björn Granvik (Prevas)

This comprehensive study investigates the detection of hate speech online using machine learning.. Through multi-label toxic comment classification, we analyzed over 200 thousand comments from the Jigsaw toxic comment competition data on Kaggle. Our goal was to create an intelligent tool to identify six different categories of hate speech. To achieve this, we developed both classical and state-of-the-art language models. Our initial approach involved a baseline model using a simple feature extraction technique and logistic regression. We then compared this model with models utilizing more elaborate feature extraction techniques in combination with recurrent neural networks and transformers. After rigorous analysis, we found that a fine-tuned transformer-based RoBERTa model performed the best, achieving a mean macro average F1-score of 0.801. This model outperforms the previous record presented by \citet{vanAken_f1score2beat}, which achieved an F1-score of 0.791. The optimal model was applied in an online visualization setting for applications to combat toxicity online.

Zoom link to presentation: https://lu-se.zoom.us/j/64919831234?pwd=TzdnQ1Y5aC9nWEFzYW1zZ280NEhEdz09

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/230601_16HosseiniFroste.pdf


 

 

 



How to get to computer science