MSc Defences
Due to COVID-19, all MSc defences from 13 - 27 March were cancelled (both days included). Instead, all projects have been graded on the basis of the written thesis.
Computer Science
Title
Modelling natural human pointing for target prediction in VR
Abstract
Pointing is a popular selection technique for VR applications, because pointing is a natural human gesture used to reference objects. Humans are inherently inaccurate when pointing, since it usually is supplemented with other means of communication. Efforts have been undertaken to correct these inaccuracies, for instance by aiding users with visual feedback using a cursor showing where they point. These methods are not always desirable nor always possible to be used, especially because such cursors counteract the natural feeling of pointing and thus break immersion. With a data collection study (𝑛 = 13) we construct a dataset, consisting of positional and orientational data of the human body while pointing. We use this dataset to build a model, that applies machine learning to the movement data such that we can predict positions of intended targets. The distance between predicted and actual positions is on average 24.42 cm. We show that the position of the target relative to the user is an important factor for the correctness of predictions. It is difficult to predict the depth of targets correctly, especially targets in front of the users. We have constructed a model that is able to describe characteristics of natural human pointing and can be used to predict desired targets and thus act as a basis for a novel selection technique. By building this model we show that we can model natural human movement and use it for input techniques.
Time and place
12 March 2020, 11:00
Room 0-11, Sigurdsgade 41, 2100 København
Supervisor
Joanna Bergström
External examiner
Morten Borup Harning
Contact
Joanna Bergström, joanna@di.ku.dk
Title
Opportunistic Gaming: How Technology Can Support Opportunistic Use of Surrounding Objects and Crafting of Objects as Part of Spontaneous Social Games
Abstract
In this Master’s thesis, I explore and define a new subgenre of gaming called opportunistic gaming. I have performed interviews with casual users of board games and tabletop role-playing hobbyists to find problems regarding playing on-the-go. I found that the participants’ main motivation for playing is to interact with other players. I also found that how the participants want to play is highly dependent on the context they play in. The participants described three potential scenarios for opportunistic gaming: public transportation, restaurants, and parties. Creating games for these scenarios requires different decisions regarding objects, social norms, movability, play stop, technology, etc. From the interviews, I extracted a formal definition of opportunistic gaming: Opportunistic gaming is social, spontaneous play that requires fewer objects than regular games and is bounded in time. I found that technology can be used to support this type of gaming; especially regarding the lack of game pieces on-the-go. I have also performed participatory design workshops, where the participants and I created three opportunistic games. Based on the interviews and workshops, I have discussed what implications the three created games have on opportunistic gaming and Human-Computer Interaction (HCI) in general. Finally, I have extracted 9 guidelines for game designers who want to create opportunistic games. These are guidelines regarding social interaction, play setting, technology, opportunistic problems, objects, and game design. Keywords: opportunistic gaming, computer augmented games (CAG), pervasive games, mixed reality games, crafting, participatory design workshops, board games, tabletop role-playing games.
Time and place
16 March 2020, 10:00
0-11, Sigurdsgade 41, 2100 København
Supervisor
Daniel Lee Ashbrook
External examiner
Nina Louise Barkhuus
Contact
Daniel Lee Ashbrook, dan@di.ku.dk
Title
Labeling Schemes for Implicit Representation of Relations
Abstract
A labeling scheme assigns a label, consisting of a bit string, to each element of a relation such that it is possible to determine if some element is in the relation by only considering its label. We consider both static and dynamic labeling schemes for relations. We present the existing work in labeling schemes for both graphs and relations. In the case of labeling schemes for general relations, we state a new, corrected version of the lower bound on the label size. Based on existing work [SG19], we define and show upper and lower bounds for the label size of relations induced from formulas of first-order logic. These results are then extended to families of relations induced from formulas of second-order logic. Moreover, with inspiration from existing work [Glo20], we define the notion of dynamic labeling schemes for relations induced from both formulas of first- and second-order logic, respectively, and state upper and lower bounds for such labeling schemes.
Time and place
16 March 2020, 15:00
S029, Building 1, Universitetsparken 5, 2100 København
Supervisor
Jakob Grue Simonsen
External Examiner
Mads Rosendahl
Contact
Jakob Grue Simonsen, simonsen@di.ku.dk
Title
Compiling a reversible cryptography language (Hermes) to ARM64
Abstract
This thesis revolves around the reversible programming language Hermes, currently being developed at DIKU as a domain-specific language for implementing symmetric cryptography algorithms on small devices. The reversibility of Hermes makes us able to give certain guarantees for its safety, in that certain sidechannel attacks such as information leakage becomes impossible. I describe these side-channels as well as reversibility and explain the intention of securing small devices. I develop a compiler from Hermes to a Reversible Single Static Assignment (RSSA) representation, which could serve as a common ancestor for multiple target assembly languages and help shorten each compilation step to a target assembly language. I also develop a compiler from RSSA to ARM64 assembly language with abstract register names. This new implementation gives more control over side-channels as we are no longer dependant on gcc and zerostack, both of which are being used in the currently existing Hermes to C compiler. I also implement the 128-bit block cipher algorithm Twofish in Hermes, thereby showing the usefulness of Hermes as a domain-specific language for implementing symmetric cryptographic algorithms. Finally, I compare my Twofish implementation with two reference implementations written in Python and C.
Time and place
23 March 2020, 13:00
A101, H.C. Ørsted Institutet, 2100 Copenhagen
Supervisor
Michael Kirkedal Thomsen
External examinor
Mads Rosendahl
Contact
Michael Kirkedal Thomsen, m.kirkedal@di.ku.dk
Title
Using Model-Based Testing to test MCP Instance-Specifications
Abstract
This thesis deals in model-based testing and model generating in the maritime sector, specifically in the Maritime Connectivity Platform (MCP). After accounting for the benefits the MCP will receive on the basis of model-based testing, a manual model-building module is presented, which introduces such functionality. The MCP has implemented service instance specifications, that are to describe service instances in the form of an xml document. These are analyzed, where after an optimized specification structure is presented, which will allow for semi-autonomous model based testing. Eventually, the behaviors of the two model generating methods are verified through a series of function- and scenariooriented tests, some of which conclude that the benefit of the automatic model creating module is near-equivalent to that of the manual module. It is concluded that the two model generators are beneficial in their separate ways, with the automatic solving the problem statement of this thesis, while the manual paves the way for future unthought of, modelbased functionality.
Time and place
23 March 2020, 14:30
A101, H.C. Ørsted Institutet, 2100 Copenhagen
Supervisor
Michael Kirkedal Thomsen
External examiner
Mads Rosendahl
Contact
Michael Kirkedal Thomsen, m.kirkedal@di.ku.dk
Title
Maritime Connectivity Platform Framework & Public Key Infrastructure Distribution
Abstract
The aim of this thesis is to develop a database server solution, which in the future could be a self su‑cient system in which the Maritime Connectivity Platform (MCP) would require none or little maintenance. This thesis will show the process of making a new server based on the current implementation of the MCP. To start out, an analysis of the current MCP server is performed to obtain an overview of how to implement a new database server, which can delegate some of the data handling away from MCP. The strengths and weaknesses of distributed and centralized server approaches will be discussed in order to determine the best cause of action. The direction chosen is a distributed database implemented in Erlang using a Mnesia database, along with gen_statem to handle the server functions. The HTTP aspect of the system is implemented using gen_tcp. The system will be tested through functional testing to ensure the functionality of the server is working as desired, along with a stress test which will be executed in order to determine its durability. The end result of the system shows that the system is able to handle all its normal calls along with handling server crashes without any issues.
Time and place
23 March 2020, 16:00
A101, H.C. Ørsted Institutet, 2100 Copenhagen
Supervisor
Michael Kirkedal Thomsen
External examiner
Mads Rosendahl
Contact
Michael Kirkedal Thomsen, m.kirkedal@di.ku.dk
Title
Automatic Image Cropping using Supervised Learning
Abstract
Image cropping when done manually is a task frequently found both time consuming and uninteresting. In Computer Vision, the task of selecting the most relevant region in an image to crop around it and produce a more aesthetic image has often been attempted with 2 approaches: Producing multiple candidate images and evaluating these or by detecting the most salient region and drawing a bounding box around it. Automatic image cropping proves difficult to both solve and evaluate because of the subjective nature of the task; there is no one right answer. Gracenote editors have for a substantial amount of time been manually cropping images for advertising use and have accumulated a catalogue of more than 50000 images with multiple crops per image with strict aspect ratios. In this work, rather than trying to evaluate an image’s aesthetic, we propose a Machine Learning approach that takes advantage of our available data to train models in a supervised manner. We present an architecture that allows us to predict multiple crops with the same model with each crop being restricted to a specific aspect ratio and show that editor behaviour can be successfully modeled through the use of a Convolutional Neural Network base and fully connected layers that perform regression. Different styles of architectures and ways of training them are compared and we achieve close-to-perfect mean performance on some aspect ratios. This contribution is extended by not only demonstrating that a new aspect ratio can be added to network with a very low data requirement through transfer learning, but also showing how modifying the standard bounding box output format can lead to increased performance and allow us to train new aspect ratios with no matching ground truth at all. Finally, we propose a method with which the model can evaluate its own crops and learn to identify when it produces a poor output and show how we, with the help of post processing steps, can catch and remedy situations that would otherwise have our crops be automatically rejected by the editors that are to receive them.
Time and place
27 March 2020, 09:00
2-0-04, Universitetsparken 1, 2100 København
Supervisor
Søren Ingvor Olsen
External examiner
Morten Pol Engell-Nørregård
Contact
Søren Ingvor Olsen, ingvor@di.ku.dk
Title
Visual tracking of rodent behaviour using deep learning
Abstract
In neuroscience rodent behavior is studied as an outcome of drug effect or disease progression. However, manual tracking of complex behaviors like rearing, grooming or interaction with objects or other animals is very time-consuming whereas visual tracking using computers is both non-biased and less time-consuming. The aim in this report is to provide a working implementation of a machine-learning algorithm, which can detect the position and posture of a rat. The postures which I intent to track is whenever the rat is rearing, grooming or neither. Herewith when the rat is in a vertical position it is measured as rearing and when the rat is sitting still or itching its stomach or ears it is measured as grooming. Furthermore, I will also evaluate the model by comparing the results with a current tracking solution used in behavioral neuroscience. To reach the goal, a database of annotated images was obtained. The database consist of more than 4000 frames with the position and posture of the rat in the frame. The database was used to train a machine learning model. To improve the model I applied an outlier algorithm to remove outliers and then a Kalman filter to estimate the values from the removed dataset. The final model can outperform the current tracking solution used in behavioral neuroscience. The model can track the position of a rat with an average error of 0.94 cm. Additionally, another model was used to predict the posture of the rat and was able to get an F1 score of 0.80.
Time and place
27 March 2020, 13:00-14:00
Room 1-1-n116b, Universitetsparken 1, 2100 København
Supervisor
Oswin Krause, Melanie Ganz, Mikael Palner
External examiner
Morten Pol Engell-Nørregård
Contact
Oswin Krause, oswin.krause@di.ku.dk
Sundhed og Informatik
Title
Min ADHD-profil - Designforslag af et støtteværktøj til elever med ADHD i en folkeskole
Abstract
In 2012, the Danish government passed a law in Denmark to cater to children with special needs in primary school. This thesis investigates whether an IT- support tool, Min ADHD-profil, can help support and solve everyday challenges for students diagnosed with ADHD in primary school. ADHD stands for Attention Deficit Hyperactivity Disorder and is characterized by lack of attention and hyperactivity. Furthermore, children diagnosed with ADHD may struggle with planning and completion of activities and tasks, as well as managing social difficulties. I have made a comprehensive literature study to gather a deeper understanding of this topic. The articles I have found through extensive research have given me an insight on what difficulties children with ADHD may have and the support tools for these difficulties. In addition, the current support technologies have been examined to see which problem areas have been covered. In this thesis, qualitative interviews have given me insight on the everyday life of a teacher and the difficulties they experience in working with students with ADHD. The collected empirical data from the initial interviews was categorized in an affinity diagram. Based on the analysis, seven important elements that are relevant to the development of a support tool for students with ADHD were clarified. It is especially important to give the student an overview of the agenda of the day through visual support and motivational rewards. It should also give them an overview of time and how to solve conflicts. These elements are used to create storyboards to illustrate the context in which the user will use the support tool. Subsequently, high- and low-fidelity prototypes have been made and have been tested on teachers, as well as on three students with ADHD. Prototyping and testing on the intended users indicate a success of the support tool, Min ADHD-profil, which gives the student a visual overview over the activities of the day, motivates the student to complete the tasks of the day, and provides support in social contexts. In order to ensure that Min ADHD- profil can solve the challenges a student with ADHD may face, a larger study is needed. This would allow for Min ADHD- profil to be used in a daily context over a longer period of time.
Time and place
17 March 2020, 11:00
1-2-0-04, Datalogisk Institut, Universitetsparken 1, 2100 København
Supervisor
Erling Carl Havn
External examinor
Keld Bødtker
Contact
Erling Carl Havn, erling@di.ku.dk
Title
Kommunernes evalueringspraksis af velfærdsteknologi
Abstract
Dette speciale er en empirisk undersøgelse af, hvordan to danske kommuner evaluerer velfærdsteknologiske projekter, samt hvad konsekvenserne er af disse udførte evalueringer. Det er et uudforsket emne, hvorfor vi finder det interessant at belyse netop dette område. Undersøgelsen er udformet som et kvalitativt casestudie, hvor der er foretaget interviews af relevante projektmedarbejdere involveret i evaluering af velfærdsteknologiske projekter samt en konsulent fra KL’s Center for Velfærdsteknologi. Som teoretisk ramme anvender vi teori, der tager udgangspunkt i forskellige evalueringsmodeller beskrevet af Hanne Kathrine Krogstrup og Peter Dahler-Larsen, samt evalueringsmodeller som er specifikt rettet mod sundhedsområdet herunder VTV, MTV og MAST. På baggrund af vores empiri, identificerer vi tre interessante temaer. Vi belyser samarbejdet mellem kommunernes velfærdsteknologiske afdeling og personalet på kommunernes pleje- og rehabiliteringscentre, når en evaluering skal foretages. Efterfølgende giver vi et dybdegående indblik i, hvorledes rammer og værktøjer anvendes til at foretage evaluering. Afslutningsvis undersøger vi, hvordan evalueringer udbredes på tværs af kommuner. I diskussionen af vores analytiske fund inddrager vi litteratur identificeret i vores litteraturgennemgang.
Time and Place
18 March 2020, 8:30-9:30
2-0-25, Biocenteret, Ole Maaløes Vej 5, 2200 København N
Supervisor
Henriette Mabeck
External examiner
Troels Sune Mønsted, DTU
Contact
Henriette Mabeck, mabeck@di.ku.dk
Title
Kommunernes evalueringspraksis af velfærdsteknologi
Abstract
Dette speciale er en empirisk undersøgelse af, hvordan to danske kommuner evaluerer velfærdsteknologiske projekter, samt hvad konsekvenserne er af disse udførte evalueringer. Det er et uudforsket emne, hvorfor vi finder det interessant at belyse netop dette område. Undersøgelsen er udformet som et kvalitativt casestudie, hvor der er foretaget interviews af relevante projektmedarbejdere involveret i evaluering af velfærdsteknologiske projekter samt en konsulent fra KL’s Center for Velfærdsteknologi. Som teoretisk ramme anvender vi teori, der tager udgangspunkt i forskellige evalueringsmodeller beskrevet af Hanne Kathrine Krogstrup og Peter Dahler-Larsen, samt evalueringsmodeller som er specifikt rettet mod sundhedsområdet herunder VTV, MTV og MAST. På baggrund af vores empiri, identificerer vi tre interessante temaer. Vi belyser samarbejdet mellem kommunernes velfærdsteknologiske afdeling og personalet på kommunernes pleje- og rehabiliteringscentre, når en evaluering skal foretages. Efterfølgende giver vi et dybdegående indblik i, hvorledes rammer og værktøjer anvendes til at foretage evaluering. Afslutningsvis undersøger vi, hvordan evalueringer udbredes på tværs af kommuner. I diskussionen af vores analytiske fund inddrager vi litteratur identificeret i vores litteraturgennemgang.
Time and Place
18 March 2020, 9:45-10:45
2-0-25, Biocenteret, Ole Maaløes Vej 5, 2200 København N
Supervisor
Henriette Mabeck
External examiner
Troels Sune Mønsted, DTU
Contact
Henriette Mabeck, mabeck@di.ku.dk
Title
Lægernes arbejdsgange og Sundhedsplatformen
Abstract
Sundhedsplatformen er en elektronisk patientjournal og er implementeret i Region Hovedstaden og i Region Sjælland til at erstatte 30 forældede systemer. Specialet fokuserer på to afdelinger i Psykiatrisk Center for Nordsjælland, hvor jeg ved hjælp af et kvalitativt casestudie undersøger hvilke forandringer Sundhedsplatformen har for lægernes arbejdsgang i de to afdelinger. Den teoretiske forståelsesramme tager afsæt i informationsøkologi og i sociotekniske studier, som benyttes til at skabe et detaljeret indblik i lægernes arbejdsgang, samt de øvrige processer, der finder sted i afdelingerne under brugen af Sundhedsplatformen. Undersøgelsen viser at lægerne har forskellige holdninger til Sundhedsplatformen og at systemet fortsat lægerne udfordringer udfordringer for lægerne i begge afdelinger. Anvendelsen af systemet afviger meget fra den enkelte læge, herunder i hvilken rækkefølge lægerne udfører arbejdet afhængig af hvor lægerne befinder sig.
Time and Place
18 March 2020, 11:00- 12:00
2-0-25, Biocenteret, Ole Maaløes Vej 5, 2200 København N
Supervisor
Henriette Mabeck
External examiner
Troels Sune Mønsted, DTU
Contact
Henriette Mabeck, mabeck@di.ku.dk
Statistics Thesis
Title
Diffusion t-means for neural networks on manifolds
Abstract
In this thesis we present a probabilistic generalization of convolution to Riemannian manifolds. The generalization is based on a diffusion t-mean known from geometric statistics. After presenting the generalization, we seek to develop a sampling scheme for estimating the diffusion t-mean. Alongside the development of the sampling scheme we introduce the elements from differential and Riemannian geometry, as well as the elements from stochastic calculus needed to understand the sampling scheme. The final sampling scheme is applicable when the manifold can be covered by a single chart, and with some additional regularity on the geometry. The method is meant for applications in geometric deep learning, and we compare the sampling scheme to an estimator of another generalization of convolution; the weighted Frechet mean, which in the context of geometric deep learning is presented in.
Time and place
23 March 2020, 14:00
Auditorium 10, H.C. Ørsted Instituttet, 2100
Supervisor
Stefan Horst Sommer
External examiner
Anders Rønn-Nielsen
Contact
Stefan Horst Sommer, sommer@di.ku.dk
IT and Cognition
Title
Computerized Cognitive Training - Enhancing Cognitive Function
Abstract
This thesis will try to investigate the area of computerized cognitive training (or ”Brain Training”) and the claimed effects on higher brain functions such as working memory and cognitive control. A review of the theory on these cognitive constructs and their relation to intelligence will be presented. A broad overview of the cognitive training literature tries to elucidate the different areas of application and controversy including the commercial sectors influence on the topic. Several researchers have tried to explain results (or lack of). Combining recent years evolution in technology and the advent of mobile devices and personal computers with an experimental examination of cognitive training on cognitive control and working memory in a normal population was carried out as the CCT study. Researching the possibility of such a study and potential gains led to informative results. Seemingly, the study showed that such an experiment was possible but showed no clear results in favor of cognitive training benefits. The limitations and future considerations are discussed.
Time and place
12 March 2020, 10:00
0-11, Sigurdsgade 41, 2100 Copenhagen
Supervisor
Joanna Bergström
External examiner
Morten Borup Harning
Contact
Joanna Bergström, joanna@di.ku.dk
Title
Political Party Affiliation Prediction on Danish Parliamentary Speeches
Abstract
We investigate the task of political party affiliation prediction on parliamentary speeches. That is, given a parliamentary speech, we predict the corresponding political party to which the speaker is affiliated. We utilize state-of-the-art neural network methods on text classification to perform predictions and evaluate performance by comparing against classical Support Vector Machines. Our primary contribution is to perform party affiliation prediction on two non-English multi-party parliamentary datasets, namely a novel (so far unpublished) Danish dataset consisting of speeches from 10 different parties, and a Norwegian (published) dataset with speeches from 11 different parties, as well as on their union, containing 21 different parties. We tune model parameters and feature configurations through 5-fold cross validation before evaluating predictions on held out test sets from each dataset. For the Danish dataset, given a specific speech, the best performing model achieves a F1-score of 0.59 and a Balanced Accuracy score of 0.57, and for the full Norwegian dataset, the best performing model achieves a F1-score of 0.56 and a Balanced Accuracy score of 0.55. These results compare favorably to prior work on party affiliation prediction on multi-party datasets. We moreover find that the state-of-the-art neural network method performs on par or better than Support Vector Machines, despite being trained on fewer words. Conclusively, we investigate and discuss implications of different pre-processing approaches as well as intricacies and ideological differences inherent in the datasets, and present results from further sub-experiments, suggesting that political parties which differ more ideologically, are more easily distinguishable.
Time and place
27 March 2020, 10:00
Supervisor
Christina Lioma & Jakob Grue Simonsen
External examiner
Peter Dolog
Contact
Christina Lioma, c.lioma@di.ku.dk