Events
CS MSc Thesis Presentation Day March 16 2022
Föreläsning
From:
2023-03-16 09:15
to
15:00
Place: See information for each presentation
Contact: birger [dot] swahn [at] cs [dot] lth [dot] se
Save event to your calendar
Five MSc theses to be presented on Thursday March 16, 2022
Thursday March 16 is a day for coordinated master thesis presentations in Computer Science at Lund University, Faculty of Engineering. Five 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:2405 (Glasburen) and E:4130 (Lucas). 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.
E:2405 (Glasburen)
09:15-10:00
Presenters: Erik Nord, Robin Rasmussen Vinterbladh
Title: Reducing costs of manual regression testing using prioritisation and partitioning techniques
Examiner: Emelie Engström
Supervisors: Per Runeson (LTH), Peter Jansson (Sony Nordic)
Regression test suites grows larger over time which leads to regression testing as an activity becoming increasingly costly. Automatic regression testing is cheap enough for high testing frequency. Manual regression testing is expensive, however equally necessary for test coverage. Using data sources already available to Sony, we introduce regression test selection (RTS) techniques for manual regression test sessions. Data sources are historic test case performance and traceability between functional requirement specifications (FRS) to test cases. We present two RTS techniques for solving the problem. One primary RTS technique selecting a sub-set of the whole test suite based on historical performance and another complementary RTS technique based on FRS partitioning. RTS techniques in this thesis were evaluated with a data simulation over stochastic data, comparative study using real-world data and continuous field study evaluating adaptations.
Link to popular science summary: To be uploaded
10:15-11:00
Presenter: Sara Hult
Title: How to develop business-critical software - a case study with a small system
Examiner: Emma Söderberg
Supervisors: Martin Höst (LTH)
The term business-critical defines systems and software whose failure may lead to loss of business or damage to reputation in the market. It can include the majority of systems and software in a business, yet the term is rarely used in literature and at companies. This thesis investigates the scope of the term, how business-critical software is developed, and what experiences can be found by implementing business-critical software techniques and methods on a small system at a company in Sweden. First, a literature study was conducted followed by a case study where three methods found in the literature study were implemented and examined on a smaller system. Thoughts and experiences were collected through interviews and a case study protocol. The result showed that there are several methods, techniques, and approaches for developing business-critical systems and software and that the three methods implemented to varying extent, all contributed to making the system safer and more reliable, however, none solved all the issues. This implies that none of the methods tried are bulletproof and that the use of several methods might be a good approach for developing business-critical software. The findings of this thesis provides insight into various methods and techniques that can be used to develop business-critical software and contributes with thoughts, experiences and opinions regarding a few of such methods, which can help others take more informed decisions and develop safer and more reliable systems.
Link to popular science summary: To be uploaded
E:4130 (Lucas)
11:15-12:00 (also via Zoom) N.B. There are no opposition places left for this presentation
Presenter: Henrik Norrman
Title: Generating abstract art using artificial neural networks
Examiner: Jacek Malec
Supervisors: Christin Lindholm (LTH), Daniel Byström (Posterton)
Artificial intelligence (AI) tools are slowly becoming integral to our everyday lives. In recent years, there has been a surge in AI projects focused on creative pursuits, such as music production, story writing, and art creation. This thesis aims to contribute to the discussion about what it can look like when AI and art meet. A two-step method is proposed—a base image generator, and an upscaling model. The base image generator is built using a generative adversarial network (GAN), trained on abstract art in the public domain. The generator is designed to generate base images, small images capturing the theme of an artist’s work.. To finalise the generation, the base image is passed to an ESRGAN model for upscaling. This method is intended as a prototype for further improvements and exploration of generative AI.
Link to Zoom presentation: https://lu-se.zoom.us/j/64451235360?pwd=S1FEK0xaSnlvd0dTZGNDT2FsQ1FHUT09
Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/230316_11Norrman.pdf
13:15-14:00 N.B. There are no opposition places left for this presentation
Presenters: Ewada Tsang, Cecilia Huang
Title: Machine Learning Based Code Generation of Security Patches
Examiner: Pierre Nugues
Supervisor: Noric Couderc (LTH), Christoph Reichenbach (LTH), Emil Wåreus (Debricked AB)
This thesis explores using CodeT5 for automating security patches through a machine learning model. It examines how dataset size, split method, number of lines changed, unique repositories, and distribution of Common Weakness Enumerations (CWE) impact model performance. The study found that increasing dataset size doesn't always improve performance, Random split outperforms Time split, and 5 Lines of change works better for Random split while 15 Lines is better for Time split. The study also found that more unique repositories can make it more difficult for the model but potentially more generalizable. Additionally, limiting the number of CWE-IDs seem to impact performance positively. The study achieved an accuracy of 40%, comparable to VulRepair's 44%, despite using a significantly smaller dataset. This study provides new insights into previously overlooked factors that impact model performance. These findings could be useful in the development of similar models.
Link to popular science summary: To be uploaded
14:15-15:00 N.B. There are no opposition places left for this presentation
Presenters: Marcus Sundell, Marlon Abeln
Title: Explainable Demand Forecasting using Causalities and Machine Learning Models
Examiner: Jacek Malec
Supervisor: Volker Krueger (LTH), Bahram Zarrin (Microsoft)
Demand forecasting is a necessary discipline to help companies predict the future demand for certain products. This thesis aims to study the difference between traditional forecasting methods, such as ARIMA, with newer methods, such as neural network approaches, together with exploring the possibilities to combine existing forecasting results with causal signals for the different approaches. The combination between existing forecasts and the causality signals is achieved by introducing an extra additive term to the examined decomposition models. The value of this term can be calculated with the help of libraries such as EconML. By utilizing this approach, the explainability of the models can still be retained. The results show that, in some cases, using this approach will yield more accurate forecasts. It is also discussed why the neural network approaches did not achieve as good results as the more traditional approaches.
Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/230316_14SundellAbeln.pdf