MSc Defences

Title
Deep Ensemble Model, as a potential solution to overcome limited amount of training data in the case of biomedical image segmentation.

Abstract
Locomotor ability is one of the most important skills possessed by mammals. The action of movement is controlled by highly complex neuronal circuits which contain interneurons and motor neurons present in the spinal cord. Interactions within those special circuits have not been studied a great deal. However, a deeper understanding of this cooperation and how it changes, could be beneficial in early diagnosis of neurodegenerative diseases such as amytropic lateral sclerosis (ALS). This project focuses on delineation of interneurons inside the mice spinal cord with the usage of deep learning, more specifically Convolutional
Neural Network (CNN). Additionally, we proposed a solution to overcome an issue of a limited amount of training, with the usage of ensemble model. Our idea of ensemble is based on cross-validation and majority voting. Training was done on three different sets of training data resulting in corresponding number of models which constituted an ensemble.
Ensemble model performed better then the baseline method, Seeded Region Growing (SRG) over the test data. Obtained F1 scores show that the proposed model is strong and efficient, which allows for fast and accurate segmentation of interneurons in a high-resolution image of sliced spinal cord.

Time and place
13 January 2021 at 2PM, online

Supervisor(s)
Raghavendra Selvan

External examiner(s)
Rasmus Reinhold Paulsen

Title
Using proximal operators for solving both rigid and deformable body contact problems

Abstract 
This thesis is concerned with the theory and practical implementation of simulators for rigid and deformable body contact problems. The main focus is on the derivation of the proximal operator model and its extension to deformable bodies. The thesis also touches on the physical laws that form the foundation of FEM based- and rigid body simulators and presents a two-phase collision detection scheme.

The derived simulators are implemented through Python and evaluated. Using this, we provide a portfolio of results showcasing the strengths and weaknesses of the proximal operator model and the gauss-seidel PROX scheme for solving it. Our results show that the implementations consistently behaved well, although complex configurations such as stacking and large mass-ratios proved difficult to handle. The extension of the PROX methods to deformable bodies is thereby shown to be feasible, but more experimentation should be put into optimizing solver parameters to make it perform well.

Time and place
6 January 2021, 10:00-14:00

Supervisor(s)
Kenny Erleben

External examiner(s)
Morten Pol Engell-Nørregård

Title
Automation of Meshing for Bio-mechanical Simulations 

Abstract
In this work, we present tools for creating background meshes with sizing functions, which can be used to define the mesh layout of a volumetric mesh. Furthermore, we offer a pipeline for finding a meshing tool’s best parameter values for different meshes, and we are using multiple quality measures to make sure that we are creating high-quality meshes.

We are using biomechanical models to showcase the difficulty in designing mesh layouts, as well as the volume meshes we can create with such mesh layouts and with our pipeline. Our results show that we can create complex mesh layouts that provide high-quality meshes, which ends up performing well in simulations.

All of the tools we have created have been placed in the easy to use, open-source library libISL. This library runs on Jupyter notebook with python and is set up such that people without much programming knowledge can use it, to create high-quality meshes.

Time and place
6 January 2021, 10:00-14:00

Supervisor(s)
Kenny Erleben

External examiner(s)
Morten Pol Engell-Nørregård

Title
Identifying Archaeological Sites in Greenland in High Resolution Satellite Images Using Deep Learning

Abstract
In Greenland alone, there are over six-thousand confirmed archaeological sites, and many of which are recognized as UNESCO World Heritages sites. It is believed that still a large number of undiscovered sites exist throughout Greenland, which are now at the risk of disappearing because of climate change.

We attempt to train a deep neural network to detect such archaeological sites in satellite images. Specifically, we present a dataset that has been specifically created for this task that consists of 247 composite RGB georeferenced images captured around the west coast of Greenland. All the images have been extracted from Google Earth Imagery. The archaeological sites in the images are primarily Thule winter settlements (1300 - 1900 CE) as they are more easily identified in the given satellite imagery than other types. There are a total 373 Thule house ruins included in the dataset.

Using this dataset, we create and train a classification model that is able to detect if there is an archaeological site present in a corresponding image. We use an EfficientNet B0 architecture pre-trained on ImageNet dataset combined with custom data augmentations. The model achieves an AUC score of 0.8508, with true positive count of 23, false positive count of 54, true negative count of 355, and false negative count of 16 on the test set.

These results demonstrate that in principle deep-learning can be used to detect archaeological sites. Given more positive examples, we believe that such a model could become a valuable tool helping archaeologists.
We extract and classify new images, illustrating how an archaeologist could incorporate our system into their workflow.

Time and place
8 January 2021, online

Supervisor(s)
Christian Igel, Rasmus Fenger-Nielsen, Rasmus Fensholt

External examiner(s)
Morten Pol Engell-Nørregård

Title
Understanding Java Vulnerabilities

Abstract
Despite its prevalence in modern society, buggy software runs far and wide, as programmers may use code containing vulnerabilities. Java’s commonly used serialization mechanism can expose a program to attacks if security is not taken into consideration. While existing countermeasures exist to address vulnerabilities ahead of time, exploits leveraging type confusion vulnerabilities may still bypass detection of static analysis tools. In this thesis we focus on CVE-2020-2805, an exploit using early referencing through serialization to achieve type confusion. We provide an in-depth analysis of the exploit, its inner workings, and how it can be leveraged to hide edges in call graphs, a feature commonly found in static analysis tools. Our analysis finds that the exploit succeeds most reliably on Java 1.8, but also works on Java 14 and that 100 versions are vulnerable. Additionally, we provide an approach to potentially detect malicious code that uses this type confusion vulnerability by flagging Method Handles whose underlying method return types may differ from the declared type.

Time and place
8 January 2021 at 10:00-11:00 AM, online

Supervisor(s)
Alexandre Bartel

External examiner(s)
Andrzej Wasowski

Title
RevCon: An Extendable Concurrent Reversible Programming Language Collection

Abstract
Most reversible programming languages are limited to non-concurrent programs. Though this is useful for demonstrating the potential of reversible computing, it hinders the real-world applicability since general-purpose computer systems never only execute a single task. In order to combat this, reversible abstractions for concurrency must be considered. The field of reversible concurrency consists mainly of theoretical work, and only a few implementations have been proposed. By extending reversible environments with concurrency, they can explore the design space of more complex structures like operating systems and solutions to problem statements related to general concurrency.

In this thesis, RevCon, a proof-of-concept programming language collection for reversible concurrent computing is presented with formal semantics and a working interpreter written in Haskell. The collection consists of a core sub-language and two language variants based upon irreversible coroutines and the actor model. The programming language framework is designed to be easily extended with new variants for other concurrency models. We show that RevCon is capable of describing and solving different problem statements, such as client-server interaction, generators of infinite sequences, and collaborative distribution of computational tasks.

Time and place
Jan 12, 2021, 12:30-14:00

Supervisor(s)
Robert Glück

External examiner(s)
Ulrik Pagh Schultz

Title
Misaligned practices: Breaking enterprise system enactment

Abstract
The purpose of adopting Enterprise systems by large organisations is to increase productivity through digital support and connecting synergistic business units. In this thesis I investigate the use of GEMS, an enterprise system in the making, used by the global organisation Falck. Applying ethnographic methods I observe and analyse the misalignment between GEMS and its the situated practices across multiple business units divided over different geographical sites. Surprisingly, I discover discover that GEMS was not used across the Falck organisation, despite being implemented 4 years ago.
Furthermore, in those 4 years the system have continuously been customised and supported by the Falck organisation. Even Most surprisingly I found that practices between business units were not influenced by each other through the connection of Gems. While Falck expected the system to standardise the processes across business units, I found that the practices were not impacted by GEMS, nor did it transform the organisation. GEMS were unable to be enacted by the organisation because of misalignment between the situated practices and the GEMS documentation practices. The contribution of this thesis is in proposing to use the structurational model of technology to explain the misaligned practices between enterprise management system and organizational practices to understand how technologies are inhibited from being enacted making it unable to influence its users.

Time and place
13 January 2021 at 12:30, online

Supervisor(s)
Pernille Bjørn

External examiner(s)
Troels Mønsted

Title
Including the Residual: A Study of Classification Work and the Gender Gap on Wikipedia via Biographies of Computer Scientists

Abstract
The Wikipedia gender gap is a multi-layered phenomenon involving the content, editors and readers of the world's largest encyclopedia. This thesis looks at its effects and related conflicts through the lens of classification and categorization and is specifically focused on biographical Wikipedia articles about computer scientists, since the field of computer science has a history of gender bias and male dominance. Contrary to previous research surrounding the Wikipedia gender gap, which is often focused on a male vs. female binary gender system, we include all gender identities in this analysis. The results indicate that classification work is a core part of Wikipedia editing and produces residual categories, which make certain groups invisible. We also gain insights into strategies used for inclusive editing, which use the existing classification system to infra-structure the aforementioned groups into the core of Wikipedia.

Time and place
14 January 2021 at 13:00

Supervisor(s)
Pernille Bjørn

External examiner(s)
Nina Louise Barkhuus

Title
Photometric stereo with refraction

Abstract
This thesis works on extending photometric stereo (PS) techniques to situations involving refractive materials. Specifically, it works on the problem where a (interface) plane is placed between camera and the surface, indicating a change in the index of refraction (IOR). It describes the optical effects this plane would have, and uses this to solve this problem for two cases, one with an orthographic camera and one for a pinhole camera projection, both calibrated and with directional lights. These models are tested on synthetic data, and found satisfactory one data with one diffuse light bounce, and up angles of around 30◦. Problems that arose when this was not the case where described and the data analysed, to create a starting point for further improvements. 

Time and place
12 January 2021 at 14:00, online

Supervisor(s)
Francois Bernard Lauze

External examiner(s)
Kim Knudsen