MSc Defences Spring 2022

Study Programme

Computer Science

Title

Detecting ships and estimating utilization factor using efficient deep learning

Abstract

The shipping business is a relatively well-established classic branch of industry, where modern advancements in computer science are yet to be utilized to a broad extent. One of such applications is an estimation of ship utilization factor based on the satellite data to provide better insights into

the demand within the transportation network. To provide meaningful estimates first, the competitor ships need to be localized on the remote sensing data and put in the context, to achieve that it is required to annotate each ship with its source, destination, etc.

Finding ships in the middle of the route demands a huge amount of processing as there is a very broad search area. MarrineTraffic registry is used to narrow down the initial search field by triangulating expected ship locations based on the reported locations. The search area is then matched against available remote sensing data to extract the possible ship sightings. LENET is used as a filtration network to discard obvious non-ship alike objects such as an empty patch of water or land. UNET is then applied to make a final ship detection on the prefiltered data.

It has proven possible to achieve ship detection with both accuracy and efficiency that could be used in practice. Yet data temporal availability is too low to create a commercially viable solution as of currently. With advancements in the field of remote sensing, in particular, cube satellites and their lunches becoming noticeably cheaper in the current time the data is bound to become significantly cheaper. The presented approach is a proof of concept that it is possible to optimize the shipping pricing further achieving better profit margins over the competitors. Obviously, the system has to be further refined and expanded before it is placed into the production environment.

Supervisor(s)

Fabian Cristian Gieseke and Klaus Kähler Holst, Maersk

External examiner(s)

Ira Assent

Time and place

21 March 2022 at 14:00 Online

Study Programme

Computer Science

Title

Not Just a "Man In a Dress": Voicing the Invisible Women of Computer Science

Abstract

Denmark and other Western Countries is experiencing a wide gender gap when it comes to Computer Science and has been so for many years creating biases and exclusions when shaping the agenda for digital development, as well as challenging the industry in dire need of skilled individuals. In an effort to contribute to changing this narrative
this thesis answers the research question: “What are the embedded stereotyped attitudes and beliefs on gender and Computer Science encountered by women students and what complexities shape the experiences of social belonging?”

To explore these embedded stereotyped beliefs and attitudes, ten former and current women students who started their Bachelor Program in the years between 2010 and 2015 were recruited for interviews. These ten women make up more than 12 % of the women who started at the Computer Science Bachelor Program during this period.
The interviews were performed as Life Story Interviews with an interview guide based on the author’s own experiences as a Computer Science student in that same period. Through coding and thorough analysis of the interviews, 5 themes emerged which has formed the basis of this study, covering significant gender stereotypes about women and attitudes towards different aspects of Computer Science as a field, as well as how
the social groups played a significant role in the women’s well-being.

What this thesis found was firstly that it is confirming that there are limited narratives about women that shape peoples experience, confirming existing research on this area, and we are extending this existing research by voicing these women’s experiences.
Secondly, we produce and propose a new concept called Relational Visibility to describe the challenge these women face, balancing how to be both visible and invisible at the same time. And finally, we propose a framework which describes the multilayered complexity which this gender minority needs to navigate when starting in Computer Science.

Supervisor(s)

Pernille Bjørn and Valeria Borsotti

External examiner(s)

Nina Louise Barkhuus

Time and place

23.03.2022 10:00 - 11:00

Sigurdsgade

Title

Images segmentation of butterfly wings using U-Net

Abstract

Throughout this thesis, we have created a data-set of shapes for the left and right fore-wings of butterflies collected by The Natural History Museum of Denmark. The dataset consists of among other things, binary image representations of wing segmentations and signed distance fields of these representations. The image segmentation of the wings was done using a UNet and the segmented wings have been transformed using signed distance function - which can hopefully be used for future research. Instead of placing hand-selected landmarks on the wings as we have exclusively focused on the whole shape. We have used image registration to look at the asymmetry between the left and right fore-wings, to i.e., assess the quality of the segmentation and study the variation between species.

Supervisor(s)

Supervisor: Mads Nielsen
Co-Supervisor: Stefan Horst Sommer

External examiner(s)

Rasmus Reinhold Paulsen

Time and place

04.04.2022 14:00

Vibenshuset

Title

RRust, a embedded reversible DSL

Abstract

This project explores using meta-programming to implement a memory safe reversible programming language. It shows that existing reversible
languages, such as Janus, can have issues with memory safety. We show
that we can solve these issues by using a memory safe language as our
intermediate language. We also show that this can be implemented with
meta-programming. We show that our new language does not have the
same memory safety issues that are in existing implementations. To show
that our language is reversible we preset a transformation of it into Janus.
We also present a transformation of the language into Rust which is our
target language.

Supervisor(s)

Michael Kirkedal Thomsen

External examiner(s)

Maja Hanne Kirkeby

Time and place

06.04.2022 11:15

HCØ, Aud 8

Title

Actively Learning to Segment With Image Loss Gradients

Abstract

Uncertainty measures in deep learning try to transparently describe the prediction quality in often opaque models. A novel way to quantify uncertainty in image segmentation is presented consisting of loss gradients with regard to the input image, referred to as image loss gradients. This measure is evaluated through a series of both qualitative and quantitative experiments showing significant correlation between large image loss gradients and wrongful segmentations. Softmax entropy is compared to image loss gradients, both of which prove to be effective uncertainty measures, but with softmax per-forming better. Image loss gradients are then tried in an active learning setup where sampling using uncertainty has often proved successful in improving model accuracy using fewer labels. Another series of experiments ultimately show encouraging signs of effectiveness that are, however, both small and statistically insignificant. Softmax entropy is again evaluated in comparison to image loss gradients, and proves even more effective in this setup.

Supervisor(s)

Jens Petersen

External examiner(s)

Rasmus Reinhold Paulsen

Time and place

06.04.2022 09:00 - Up1-2-0-04 

Online

Study Programme

Computer Science

Title

Closing the loop: Effective affect modelling and adjustment in a live first-person horror game environment

Abstract

Recent years have seen a lot of research output produced in the field of player modeling. Approaches may vary, but the common goal remains the same: to effectively model a player’s personal experiences.
Coupled with increasing computational capabilities and quickly advancing research in applying AI to player modeling, we now face a very exciting possibility: not only can we design and cater a video game to a specific audience, but also we can develop a fully interactive experience that will live and evolve, depending on a single player’s input and perceived emotion. We could go as far as to say that we’re not really designing a game as much as a template for one, that adapts and moulds itself to each
player.
This thesis presents how state-of-the-art findings in the field of affective computing can be used to develop an immersive first-person horror game experience. Our goal is to present a comprehensive overview of developing an ”AI Director” - a system capable of:
1. Estimating a player’s current emotional response (more specifically: their arousal)
2. Selecting a subset of available in-game actions to steer the response in the desired direction
3. Detecting whether the selected course of action yielded expected results
a) In case the result was different from expectations - perform necessary adjustments to account for that discrepancy in future iterations.

The idea of player modeling is nothing new - countless games gather and process gameplay data to estimate engagement rates and there have been a few titles where ”Artificial Intelligence” systems are used to moderate the player’s individual experience. We argue, however, that none of our predecessors effectively combine the state-of-the-art methodologies presented in this thesis - unbounded and relative affect annotation for measuring affect in data gathering; leveraging the ordinal annotation for
better in-training results in our Neural Network; applying the Learning Under Privileged Information paradigm for affect modeling with physiological inputs serving as privileged data. We combine all of the state-of-the-art technologies into a robust affective system situated at the core of our horror game. We then evaluate the resulting end-to-end solution by performing a user study on test subjects not previously observed in our data.

We introduce a typical attempt at building an affect recognition system - we conduct a data collection experiment, where subjects are asked to play the game and self-report their perceived affect afterward. Using this information, we fit an inference model which is capable of estimating the player’s emotional state. The novelty of our approach lies in using a relatively new machine learning paradigm - Learning Under Privileged Information (LUPI) - to be able to use physiological sensory data, unavailable at the time of inference, to train the model. We further develop a system capable of running in a self-contained loop, where the player’s emotions are evaluated and then steered towards a desired response. The system then measures the expected vs. actual response and corrects itself to achieve maximum accuracy. The final results are be evaluated by performing a comparative user study of participants playing both the affective and regular variant of the game.

At the time of writing this thesis, we believe that none such system has been attempted before. The novelty of our work lies in incorporating the aforementioned cutting-edge technology into a fully developed game, which was built as an immersive, adaptive experience. Most of the research mentioned and used by us throughout this thesis has only been tested under clinical conditions, with no practical applications. There have been notable attempts to create games which incorporate physiological feedback from the player to modulate the experience (Nogueira et al. 2016) but none of them make use of the LUPI paradigm and therefore require the players to wear some sort of a sensing device. What is more, we could not find any implementations of a full-feedback loop in those
games. They mostly predict the users response and then just progress through a set of pre-determined affective states - more often than not assuming that the player’s fear levels can only rise. In our work, we introduce an attempt at a self-correcting system - capable of evaluating its own effectiveness and adjusting suitable parameters.

We produced a number of useful results in our study - starting with the evaluation of unbounded affect annotation as a way of gathering arousal ground truth data for video game playthroughs. The annotation protocol proved to be very accurate, with a number of physiological features showing high degrees of correlation with the annotated value, as well as a number of features derived from it. Additionally, we have shown the tonic driver of EDA signal to be a good predictor of user’s perceived arousal - as opposed to the more widely used phasic one. We have achieved high accuracy scores from our LUPI models: using the pixel-based classifier we have achieved a 71% accuracy score on binary classification and 60% for 3 labels. Using an SVM+ implementation we have reported scores of 80% for 2 classes, 74% for 3 classes, and 66% for 5 classes. We have built and ”AI Director” system around that inference model, designed to modulate the player’s arousal to follow the desired curve. Our final evaluation has shown that system to be effective, making for a better and more scary experience.

Supervisor

Valkyrie Arline Savage

External examiner

Andrea Corradini

Time and place

25 April 2022

Sigurdsgade 41, CPH N

Room 2-03

Study Programme

IT and Cognition

Title

Slippery Slope - Exploring the Language Used To Describe Complex Friction Patterns on Touch Surfaces

Abstract

With the advent of variable friction feedback on touch surfaces, the question arises how individuals describe complex friction patterns in their own language. We present an experimental study with 24 participants to analyze the language used to describe complex variable friction patterns on touch surfaces. The study followed an semi-structured think-aloud protocol that stimulated the participants to describe their tactile experience with regard to sensory, emotional, and associative elements inspired by previous work from the vibrotactile domain [35, 36]. We found the language to be complex in relation to sensory and emotional aspects and surprisingly rich in relation to associative contents. Furthermore, we tested the application of natural language processing techniques to analyze the participants’ verbal descriptions. We generated semantic themes using hierarchical clustering on static word embeddings and evaluated these based on the three facets. The semantic model partially captured perceptual aspects evident in the participants’ verbalizations. At last, we presented a methodology to vectorize
the descriptions using the semantic concepts inspired by Girju and Peng [14]. We performed a Pearson correlation analysis on the resulting linguistic features paired with self-engineered signal features of the friction patterns in use and found several significant correlations. Low energy (RMS) signals are described as ‘smooth’ and ‘soft’, while ‘electric’ and ‘electrical’ are more frequently associated with friction patterns with high energy. We also found the pulse structure to correlate with the use of
the descriptors ‘pleasant’, ‘light’, and ‘subtle’. We consider the results of this exploratory analysis tentative. Nonetheless, we motivate future work to test the proposed methodology further.

Keywords: surface haptics • electrovibration • human perception • natural language processing • word embeddings • human-computer interaction

Supervisor

Hasti Seifi

External examiner

Eve Hoggan Christensen

Time and place

14:00 - 15:30

Online

Study programme

Computer Science

Title

End-point Projections from DCR Choreographies to Jolie Interfaces

Abstract

As distributed systems are becoming more common, and more complex, better tools for architecting and implementing them are in demand. In this thesis I aim to support the process of developing distributed systems by implementing a compiler from choreographies described in the notation of Dynamic Condition Response graphs, to DCR end-points, to interfaces in Jolie, a language designed specifically for making distributed systems. No solution for this problem exists yet. A partial solution has been developed in the wsdl2jolie project, but this does not cover the entire process from choreography to implentation, but only from end-points defined in WSDL to Jolie Interfaces. My implementation parses DCR choreographies, project the end-points of roles of the choreographies, and translate these end-points into Jolie interfaces, making a complete template for an implementation of the distributed system described by the choreography.
I have made a small unit test-suite, run examples of complete choreographies, and verified that the resulting output matches the expected interfaces.

Supervisor

Thomas Troels Hildebrandt

External examiner

Marco Carbone

Time and place

To be determined, contact Thomas Troels Hildebrandt (hilde@di.ku.dk) for more information.