Faculty of Engineering, LTH

Denna sida på svenska This page in English

Software@LTH events

CS MSc Thesis Presentations 20 June 2022


From: 2022-06-20 10:15 to 17:00
Place: E:4130 (Lucas) and E:2405 (Glasburen)
Contact: birger [dot] swahn [at] cs [dot] lth [dot] se
Save event to your calendar

Six Computer Science MSc theses to be presented on 20 June

Monday, 20 June there will be six master thesis presentations in Computer Science at Lund University, Faculty of Engineering.

The presentations will take place in E:4130 (Lucas) and E:2405 (Glasburen). See time and location for each presentation below.

Note to potential opponents: Register as an opponent to the presentation of your choice by sending an email to the examiner for that presentation ( 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.

10:15-11:00 in E:4130 (Lucas)

Presenters: Tove Sölve, Lucy Albinsson
Title: Investigating Hybrid Approaches for POI Name Matching
Examiner: Jacek Malec
Supervisors: Dennis Medved (LTH), Hampus Londögård (AFRY)

Determining whether or not two Points of Interest are the same entity, can be a difficult task due to inconsistent information between different data sources. This can cause problems such as duplicated or removed Points of Interest in map services. This thesis investigates hybrid approaches for matching Points of Interest based on their names and coordinates. We present hybrid approaches based on state-of-the-art similarity functions, semantics and machine learning. The highest performance was achieved using a random forest classifier with features based on hybrid similarity functions, with an F1-score of 0.986 and error reduction of 67% compared to state-of-the-art approaches.

Link to popular science summary:ölve.pdf

11:15-12:00 in E:4130 (Lucas)

Presenter: Dennis Jusufovic
Title: Optimizing active alignment during assembly using neural networks
Examiner: Jacek Malec
Supervisors: Dennis Medved (LTH), Martin Hellaeus (Axis Communications), Stefan Plogmaker (Axis Communications)

When a camera module is assembled a crucial step is the joining of the lens and sensor. If the sensor is not at the correct distance or angle in relation to the lens the module will perform badly e.g., bad focus. This is a complex problem since the optimal distance or angle can vary from lens to lens. At Axis the problem is solved with active alignment. Each lens and sensor pair are analyzed until the most optimal distance and angle is found, resulting in saved costs and increased production yield. In this study we try to optimize Axis’ active alignment process using artificial neural networks. We will explore three different model candidates and compare their performance. Our experiments showed that image analysis using neural networks would be a good replacement for the current approach, but that more work should be done on making the model more generalizable.

Link to popular science summary: To be updated

13:15-14:00 in E:4130 (Lucas)

Presenter: Robin Göransson
Title: Deep Reinforcement Learning with Active Vision on Atari Environments
Examiner: Jacek Malec
Supervisor: Volker Krueger (LTH)

Reinforcement learning (RL) algorithms have throughout the years found most success on artificial domains where the state-space is fully observable. Atari 2600 games, where the full screen is used as the input state, are thus commonly used to evaluate the performance of an RL agent. Using the full screen as input could however be unnecessary as a significant amount of pixels on the screen do not contain any relevant information. In this thesis a current RL algorithm is expanded using active vision. By restricting the visible portion of the screen and giving the agent means to control this vision window the state-space is made partially observable. With the addition of active vision an agent must simultaneously learn to play the game and learn to control its vision. This more challenging task is solved using a modified version of the recurrent A3C-LSTM network which can handle active vision. Throughout the thesis different models, that simulate vision in different ways, were used. While all models used rectangular focal areas the models differed from each other in how resolution was handled. The first models used a single, constant resolution in the full focal area. Next, the resolution was set to decrease with increased distance to the center of the focal area. As a final addition peripheral vision was added to the models. The peripheral vision was created using a very low resolution background outside the focal area. The addition of the peripheral did improve the performance of the models on Pong and Breakout while the addition barely affected the performance on Beam Rider. The model using decreasing resolution with the added peripheral was the best performing model on both Pong and Breakout while the most successful constant resolution model without peripheral achieved the best performance on Beam Rider.

Link to popular science summary:öransson.pdf

14:15-15:00 in E:4130 (Lucas) N.B. This presentation has now been added

Presenter: Eric Sporre
Title: Row vs. column data layout in a graph database query engine
Examiner: Jacek Malec
Supervisors: Luigi Nardi (LTH), Henrik Nyman (Neo4j)

This thesis aims to examine if there is any performance improvement to be gained by changing the memory layout from row-wise to column-wise inside of the Neo4j query engine. In order to test this a column-wise representation was created along with new implementation for a few operators to better leverage the potential of the new memory layout, such as using SIMD. This change means that the query execution strategy is changed from the current, which relies upon fusing and compilation, to vectorized approach. The conclusions drawn were that a performance improvement was achievable by combining the new column-wise layout in combination with vectorized solutions. These solutions are limited however, since they can only be used for value types and might not be suitable for all operators. The memory layout change or the use of the new vectorized implementations are not enough on their own to yield an improvement, only in combination do they improve upon the state-of-the-art compilation strategy currently in use.

Link to popular science summary:

14:15-15:00 in E:2405 (Glasburen) N.B. This presentation has now been added

Presenters: Stefan Jonsson, Emma Grampp​​​​​​​
Title: Designing a Domain Specific Language for Robotics​​​​​​​
Examiner: Jesper Öqvist​​​​​​​
Supervisor: Christoph Reichenbach (LTH)

As industrial robots get cheaper it becomes more important that they are also easy to program. For this purpose we have created a domain specific language that is used to describe movements and interactions of robots with multiple moveable parts. The language is inspired by an earlier language by M. Stenmark, but adds abstraction over components and procedures. In order to evaluate the language we have tried to translate a collection of programs from Stenmarks language to our language for comparison. We have also had a roboticist solve a collection of problems with the language and fill in a form about the experience. The language has some limitations (the main one being the lack of graphical user interface) but we hope that it can serve as a basis for further development and research.

Link to popular science summary:

15:15-17:00 in E:4130 (Lucas) N.B. This presentation has now been added

Presenters: Jonas Boström, Viktor Olsson​​​​​​​
Title: Workload detection and Continuous Automatic Bayesian Optimization in Database Management Systems​​​​​​​
Examiner: Christoph Reichenbach
Supervisors: Luigi Nardi (LTH)

The goal of this thesis has been to investigate the possibility of multi-workload optimization in Database Management Systems and workload detection. A system was successfully constructed to allow for multi-workload testing and data aggregation. The performance gain when optimizing using this project did not seem to match the optimizing performance obtained during testing within the single-workload framework. To test workload detection, data was collected for the benchmarks TPC-C, CH-benchmark and Wikipedia for two different types of metrics. The first was hardware-based metrics which was tested using the change detection technique CUSUM. It was found that hardware-metrics excelled in separating data for the chosen workloads in non-optimizing circumstances, and in optimizing situations it was found to be too unreliable. The second type consisted of the query types that were executed by the Database Management System. When tested with the DBSCAN clustering method all data points were clustered correctly.

Link to popular science summary: To be updated