Four MSc theses to be presented on March 17, 2017.
March 17 is the day for coordinated master thesis presentations in Computer Science at Lund University, Faculty of Engineering. Four MSc theses will be presented.
Note to potential opponents: Register as opponent to the presentation of your choice by sending an email to the examiner for that presentation (firstname.lastname@example.org). 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! Further instructions are found on this page.
|PRESENTERS||Jonathan Lundholm, Paul Steneram Bibby|
|TITLE||False Alarm Filtering with an External Classification Service|
|SUPERVISORS||Jörn Janneck (LTH), Lars Larsson (Axis)|
|ABSTRACT||Within surveillance, a camera's inability to distinguish between human and non-human objects often gives rise to false alarms. False alarms are alarms triggered by non-human activity, such as a tree blowing in the wind, and are usually not of interest for the operators. This thesis shows that the use of an external classification service, which can classifying objects in images, can greatly reduce the number of false alarms. A classification service was developed as a distributed system, capable of classifying images on request. The service is deployed as a cluster on a number of AWS EC2 instances. The thesis has also looked at ways to reduce the network traffic, and showed it is possible to reduce the data needed to classify images, without losing precision.|
|TITLE||Maintaining Source Origin in a Modelica Compiler|
|SUPERVISORS||Niklas Fors (LTH), Jon Sten (Modelon AB), Jonathan Kämpe (Modelon AB)|
Modelica is a modelling language used to describe a physical system and can be used for simulation of the system. A Modelica compiler takes a Modelica model and transforms it into a equation system in order to simulate it. During this transformation the information about the source of the equations are lost. This information could be useful to have when, e.g., trying to locate an error.
|TITLE||Measuring Semantic Distances between Software Artifacts to Consolidate Issues from the Development and the Field|
|SUPERVISORS||Elizabeth Bjarnason (LTH), Markus Borg (SICS)|
|ABSTRACT||Identifying and tracking different structural representations of functionally overlapping issues is important in order to keep a well-maintained issue management corpus, establishing efficient response to develop and deploy repairs to the software defects.|
While this normally is achieved with time-costly reviewing processes by special teams, this project intends allowing review-teams making better and faster qualitative assessments by providing a quantitative metric by implementing a tool using a search engine as information retrieval mechanism.
The project is a case-study at an unnamed company. We use informal interviews to define the semantic distance, but to measure and verify the accuracy of it we use qualitative and quantitative methods. As a result of the limited scope of this thesis, the tool has limited use in a live development environment. We conclude that this approach is potent and could bring fruitful findings in the issue management and issue maintenance field if developed further upon.
|TITLE||Crowdsourcing Component Selection by Text Mining Stack Overflow Discussions - Experiences from Active Learning and Self-training|
|SUPERVISORS||Elizabeth Bjarnason (LTH), Markus Borg (SICS), Rasmus Ros (LTH)|
|ABSTRACT||Component-based system engineering facilitates high-quality system development, however, the outcome depends on well-informed component choices. The objective of the thesis is to crowdsource Stack Overflow for component experiences, intended for component decision support.|
Questions and answers on the domain were classified using an SVM, and two main annotators carried out Active Learning-iterations according to a set of criteria. Finally, we investigated the potential of self-training, especially in combination with Active Learning. We evaluated annotation agreement using pair-annotation and calculated Krippendorff’s $\alpha$, which proved poor. Two common disagreement cases could be inferred, and we encourage recurrent pair-annotation to confirm consensus. Although Active Learning was evaluated on the annotators’ training sets separately, the learning curve declines slightly, and we speculate that it risk introducing randomness in the training set by borderline cases. Self-training, however, improved accuracy if used with carefully tuned parameters, and appears especially promising in combination with Active Learning.