MSc Defences Winter 2022/2023

Study Programme  

Computer Science

Title  

TRAZE: Total Resource Accounting for Zero-Emission

Abstract  

As the economy continues to grow, the demand for food, energy and materials follows suit. Climate change, biodiversity loss and pollution are globally pressing matters that require a fundamental transformation of resource-producing, consuming and recovering systems, to which sustainable supply chain management (SCM) and Carbon Credits (CCs) have gained momentum.

This thesis introduces TRAZE, a generic prototypic framework for total resource accounting of supply chain networks (SCNs), particularly addressing carbon accounting and CCs. TRAZE implements REALISTIC, an event-based modelling framework by Bager, Düdder, Henglein, Hébert, & Wu, 2022, and extends the algebraic resource accounting framework by Hébert, 2020 and the prototype by Alstrup & Borgert, 2022.

We demonstrate that it is formally possible to model and compute the carbon effect based on event data, where event data comprises assertions that something has happened or is measured. We provide a tool for interpreting data and deriving information related to SCNs and carbon emissions, and we address and analyse common attack vectors and challenges related to CCs. To prove our hypothesis,
we implement a model and compute the carbon effect applied to a simple yet illustrative use case based on the Columbian coffee SCN. In conclusion, we argue that TRAZE is a flexible framework for fine-grained tracking of production and transportation of physical resources and reliably modelling their direct and indirect carbon effects across mutually distrusting actors in SCNs.

Supervisor(s)  

Fritz Henglein
Co-Supervisor: Christina Singh

External examiner(s)  

Mads Rosendahl

Date and time  

23.12.2022 10:00-12:00

Room  

UP5, 772-01-0-S29

Study Programme  

Computer Science

Title  

eBPF and PCC

Abstract  

eBPF is a subsystem of the Linux kernel that allows for loading programs from user space and into the Linux kernel during runtime, to be executed in kernel space. eBPF is rapidly being adopted and used. There is however a severe security risk connected with running essentially untrusted programs in the kernel - a risk that is currently mitigated, but not nullified, by the in-kernel eBPF verifier. Proof Carrying Code is a concept that
shifts the responsibility of guaranteeing the safety of a program from the code consumer to the code producer, by specifying one or more safety policies by which a program must abide and requiring the code producer to
provide a proof alongside their program, that proves the program does not violate the safety policies. In this report I present Featherweight eBPF, a small IMP-like language representing a subset of eBPF, a set of safety
policies for it, a weakest precondition verification condition generator for Featherweight eBPF programs and an accompanying proof of concept proof carrying code architecture. Finally I conduct a series of experiments
comparing the capabilities of my implementation to that of the in-kernel eBPF verifier

Supervisor(s)  

Ken Friis Larsen

External examiner(s)  

Philippe Bonnet

Date and time  

10.01.2023 14:00

Room  

UP1, room 3.2.07

Name of student(s)  

Konstantinos Poul Papanikos

Study Programme  

Computer Science

Title  

Analysis of the Carbon Footprint of Sequence Aware Recommender Systems

Abstract  

Recommender systems (RS) have become ubiquitous in various domains, including e-commerce, music, social media, and more, by providing personalized recommendations based on user interactions. In recent years, there has been a surge of interest in sequence-aware RS, which make recommendations based on the sequential logs of recorded user interactions. However, the training and use of these models require a significant amount of energy, contributing to their carbon footprint. In this
dissertation, we aim to bridge the gap in the literature on the carbon footprint of RS by benchmarking five sequence-aware RS on four different datasets. Our results show that simple techniques like nearest neighbor are more efficient and perform comparably or even better than neural network models. Additionally, we investigated the model’s relationship between dataset size and carbon emission, finding that the tradeoff between the two factors depends on the specific model’s architecture being used. This suggests that different models may have different
scalability characteristics in terms of their energy efficiency.

Supervisor(s)  

Maria Maistro and Christina Lioma

External examiner(s)  

Panagiotis Karras

Date and time  

23.01.2023 10:00

Room  

1-0-04 at DIKU

Name of student(s)  

Laura Høyer Boesen

Study Programme  

Computer Science

Title  

Information Retrieval, But at What CO2st? A Simple Sample Complexity Analysis of IR Models with the Goal of Decreasing Carbon Emissions

Abstract  

The current state-of-the-art methods in machine learning (ML) and natural language processing (NLP) are highly energy-intensive, leading to a focus on green ML research that is more energy-conscious in its experimentation. In a recent study, Scells et al (2022) [2] compared the power usage and emissions of different information retrieval (IR)
methods using standard datasets, but did not examine how the choice of dataset and its structure affects the energy cost of IR. This leaves open the question of how effectiveness, energy cost, and dataset structure are related. This thesis aims to investigate ways to reduce the sample complexity of three IR models - uniCOIL, SPLADE, and DPR - with
the goal of decreasing carbon emissions. By analysing the factors that contribute to the sample complexity of IR models, strategies for reducing the amount of training data and energy resources required for these models to achieve a desired level of performance can be identified. This simple approach can help reduce the carbon footprint of IR models while still allowing them to maintain their effectiveness. The experiments show that by finding optimal sample complexity effectiveness is increased while the carbon emission impact of training decreases. Additionally, I explore the trade-off between effectiveness and carbon emission.

Supervisor(s)  

Maria Maistro and Christina Lioma

External examiner(s)  

Panagiotis Karras

Date and time  

23.01.2023 11:00

Room  

1-0-04 at DIKU

Name of student(s)  

Sebastian Holmby Hansen

Study Programme  

Computer Science

Title  

Automata synchronization in geometric settings

Abstract  

This thesis investigates the task of gathering swarms of very simple robots in geometric settings. It takes inspiration both from the much-studied field of automata synchronization, where one aims to find words that reset a given deterministic finite automaton, and the newer field of swarm robotics, where one aims to manipulate swarms of such simple robots. In the studied problem, the goal is to find sequences of movements that gather a number of robots in a single location inside a given environment.
We first define both the model and problem in geometric settings, before reviewing relevant literature and discussing the overlap between automata synchronization and swarm robotics. In our proposed version of the problem, robots are assumed to be manipulated solely by some external power that moves all robots uniformly. Next, we consider a selection of environments with varying assumptions, namely disks, polygons and curves. As the main contribution of this thesis, we introduce a number
of algorithms which solve the problem of synchronization in each of these types of environments.
Among other results, we present an algorithm with synchronizes simple polygons in time O(kαD), and asymptotically optimal algorithms for unit disks and regular polygons. Finally, we analyze all the proposed algorithms in terms of their efficiency and computational complexity, and discuss possible future directions of research.

Supervisor(s)  

Mikkel Abrahamsen

External examiner(s)  

Philip Bille

Date and time  

27.01.2023 09:15 - 10:15

Room  

HCØ Aud. 7

Name of student(s)  

Lulu Omar Aden og Emilie Louise Mossin

Study Programme  

Sundhed og Informatik

Title  

Kvalitativ undersøgelse af anvendelsen af VAR Healthcare i en kommunal hjemmepleje

Abstract  

Background: In recent public debate there have been opposing opinions about the value of procedures in the healthcare system. Current legislation demands that healthcare workers must keep up to date with the latest research, but in practice this is rarely the case. One example of this is in Municipality X, here data shows a decline in the use of the evidence-based procedure database VAR Healthcare from 2020 to 2021. This suggests that there are barriers to the use of this technology.
Purpose: The purpose of this project is to shed light on how VAR Healthcare is used within a municipal homecare service, how the technology supports the healthcare workers workflow, and why
there are barriers to its usage.
Research design: A qualitative study is conducted using field observations and semistructured interviews along with a literature study. Data from the qualitative study is analysed through an Abductive Grounded Theory approach. The theoretical framework for understanding the results is a
sociotechnical approach along with the theory of information ecologies. Findings from the literature study are used along with the theories to understand the results from our qualitative study.
Results: The project data are based on nine individual interviews, one two-person interview, field notes and results from ten existing studies. Our analysis identifies four themes that have an impact on the use of VAR Healthcare in the homecare service: Usage, Culture, Technology, as well as
Leadership and implementation.
Conclusion: VAR Healthcare is used differently by the different healthcare workers depending on their profession, but overall the technology is not widely used. The structure of the procedures in VAR Healthcare is aimed at institutions in the primary sector, making it difficult for the workers in
the homecare service to use it. The culture in the workplace is characterised by routines and habits, which is an additional hindrance to the use of the technology. These barriers suggest that VAR
Healthcare has not been implemented properly. The workers point at the managers having an important role in this implementation, however, the managers do not perceive the use of the technology as their responsibility. This suggests that both the managers and the workers are not taking responsibility for implementing VAR Healthcare in their workflow.

Supervisor(s)  

Henriette Mabeck

External examiner(s)  

Yutaka Yoshinaka

Date and time  

13.01.2023 08:00 - 10:15

Room  

Lokale 2-0-25 på Biocenteret

Name of student(s)  

Huma Muneeb Idris

Study Programme  

Sundhed og Informatik

Title  

Patientpleje og vurderingsskemaer - En kvalitativ undersøgelse af sygeplejerskers udfordringer, erfaringer og strategier ved brugen af vurderingsskemaerne

Abstract  

This paper examines the nurses at Ward X at Rigshospitalet’s daily work, when dealing with the evaluation forms from the Health Platform. The paper outlines different theories such as Abductive grounded theory, hermeneutic spiral, and actor-network theory. The applied methods and how they have been implemented in the paper is also described. The main method is based on qualitative methods, which has been used to elucidate the different aspects of the case. This has all been necessary to answer paper’s problem definition, which sounds:

Does the nurses on Ward X at Rigshospitalet experience challenges when dealing with the assessment form from the Health Platform in their daily work? If so, which ones? And do the nurses use different workarounds to accommodate any challenges, if this is the case, which ones are used? And do the challenges have any significance for the treatment of the patients?

This paper includes an analysis, which is based on qualitative interviews of the nurses at the specific ward and different research papers. The Paper also includes a discussion that puts the different statement against each other and the used literature.

Based on the paper it can be concluded that the nurses on Ward X at Rigshospitalet face a number of challenges when using evaluation forms, including difficulties in navigating the system, limitations in inputting values into the system. In response to these challenges, it can be concluded that the nurses often resort to workarounds such as documenting observations by hand and creating their own forms for documentation. Lastly it can be concluded that these workarounds may have an impact on patient care.

Supervisor(s)  

Henriette Mabeck

External examiner(s)  

Yutaka Yoshinaka

Date and time  

13.01.2023 11:45 - 12:45

Room  

Lokale 2-0-25 på Biocenteret

Name of student(s)  

Frederik Weber Wellendorf

Study Programme  

Mathematics

Title  

Fitting Distances by Tree Metrics Minimizing the Total Error within a Constant Factor

Abstract  

In this thesis we will take on the L1-fitting tree metrics problem.
We will show that this problem can be solved using a constant-factor approximation algorithm.
In order to solve it we reduce it to the Hierarchical Correlation Clustering problem, which will be the main focus of the project.
The thesis is based on the paper by Das, Kipouridis, Parotsidis, Cohen-Addad and Thorup [4]. We will review the relevant theory and build upon their algorithms and methods in order to improve the theoretical foundation.
Thus, we will show that Hierarchical Correlation Clustering can be solved using a constantfactor approximation algorithm, which will ultimately allow us to find an explicit constant for the approximation ratio, which has not been done before.

Supervisor(s)  

Mikkel Thorup

External examiner(s)  

Jørgen Bang-Jensen

Date and time  

20.01.2023 14:45

Room  

1-0-10 på DIKU

Name of student(s)  

Christian Jonas Peifer

Study Programme  

Statistics

Title  

Diffeomorphic Image Registration

Abstract  

This thesis presents an exposition of the mathematical foundations of diffeomorphic image registration and discusses how stochasticity can be introduced to this framework.
It is divided into three parts: In the first part we consider the geometry of an abstract registration problem defined by the diffeomorphic action on a set of shape objects. In this general setting the momentum and Euler-Poincaré evolution equations are derived, and these are then specialized to image registration. The second part concerns the analytical
foundations of the theory, developing the framework of admissible Hilbert spaces and their induced groups of diffeomorphisms, culminating in the theoretical guarantee of solutions to the image registration problem. Finally the third part develops the theory of Itô integrals and stochastic differential equations and we use this to discuss a recent approach on how these can be introduced to the registration problem in a natural way
preserving the geometric structure of the problem.

Supervisor(s)  

Stefan Sommer

External examiner(s)  

Steen Markvorsen

Date and time  

24.01.2023 13:00

Room  

Vibenshuset, 4. sal - mødelokalet