MSc Defences Summer 2023

Name of student(s)  

Lennard Reese

Study Programme  

Computer Science

Title  

Efficient streaming algorithms for metric first-order temporal logic in Timely Dataflow

Abstract  

Abstract
As computer systems become increasingly interconnected and digitalization becomes more
prevalent, the volume of data to be stored, processed, and analyzed continues to grow. Ensuring the correctness and security of computer systems and their outputs is of paramount
importance, requiring the need for real-time monitoring and runtime verification.
This thesis focuses on extending an existing monitoring tool that analyzes logs based on a
predefined security policy and reports any policy violations, enabling appropriate actions to be
taken. Specifically, our work involves the development of efficient streaming algorithms for temporal logic. These algorithms are designed to be highly parallelizable and capable of handling
out-of-order data, allowing for real-time monitoring and verification of system behavior. The
algorithms are implemented in the Timely Dataflow framework.
The results of our research demonstrate the effectiveness and efficiency of the developed
streaming algorithms in handling temporal logic operations on real-time data streams. Through
evaluations and comparisons with existing tools, we show that our algorithms provide scalable
and accurate monitoring verdicts, enabling the detection and reporting of policy violations in a timely manner.

Supervisor(s)  

Dmitriy Traytel

External examiner(s)  

Alceste Scalas

Date and time  

12-06-2023 at 10:00

Room  

Sigurdsgade 41, room 2.03

Name of student(s)  

Jens Tobias Dahl Larsen

Study Programme  

Computer Science

Title  

Database-style Indices in a Formally Verified Runtime Monitor

Abstract  

Abstract
Runtime monitors are tools for correctly detecting violations of a specification. The correctness of the algorithms of such monitors, can be validated
through formal verification. VeriMon is such a monitor, formally verified in
the proof assistant Isabelle. It solves the monitoring problem through operations on finite tables, but it employs naive underlying implementations for
these. In this thesis, we equip the tables of VeriMon with database-styled
indices, which can be thought of as mappings from a subset of a table’s
columns to the remaining columns. We provide a nice interface for users to
specify these particular columns. We optimize table operations through the
indices, in particular the natural join operation. Additionally, the correctness
statement of VeriMon is preserved, by relations, linking our extension to the
original implementation.

Supervisor(s)  

Dmitriy Traytel

External examiner(s)  

Alceste Scalas

Date and time  

12-06-2023 at 11:00

Room  

Sigurdsgade 41, room 2.03

Name of student(s)  

Xiaohan Wang

Study Programme  

Computer Science

Title  

A study of information retrieval task using BERT embedding

Abstract  

Abstract
In recent years, as computing power has increased and more and more models
have been proposed, neural networks have become more widely and effectively
used in various fields, and information retrieval tasks are no exception.
In order to ensure that the neural network model can effectively process the
input text, the first task is to vectorize the text and get word embeddings.
The quality of the word embeddings are important to the subsequent processing, as high-quality word embeddings lead to better results.
In 2019, BERT[1]
(Bidirectional Encoder Representations from Transformers)
was proposed, with its large-scale pre-training process to learn the contextual representation of text, allowing us to obtain high quality word embeddings. After this, many researchers combine the word embeddings obtained
by the BERT model with other models, such as CEDR[2]
(Contextualized
Embeddings for Document Ranking), which combines the word embeddings
obtained by the BERT model with the KNRM[3]
(kernel based neural model)
model and obtains good performance results.
Our main task in this thesis is to reproduce and evaluate some of the results
of CEDR on the Robust04 dataset. In addition, we perform significance tests
and case studies for the results and conduct new experiments to improve the
performance of the model by adding valid information and interpolation.

Supervisor(s)  

Ingemar Johansson Cox

External examiner(s)  

Peter Dolog

Date and time  

13-06-2023 at 10:30

Room  

Online

Name of student(s)  

Martin Max Kristensen and Jonas Juul Hansen

Study Programme  

Computer Science

Title  

A Novel Approach to Euclidean Hub Labelling

Abstract  

Abstract
In this thesis we study the Euclidean Shortest Path Problem (ESPP), the problem of
finding a shortest path between two points in the plane among a set of polygonal obstacles. Recently, Du et al. (2023) proposed an algorithm based on hub labelling where,
in a preprocessing step, a uniform grid is superimposed on the plane in which each cell
contains the hub labels from obstacle vertices with visibility to it. We propose an alternative planar subdivision that we call a visibility subdivision, which has the desirable
properties that (1) all points in the same region can see the same set of obstacle vertices
and (2) no other subdivision satisfying (1) has fewer regions. In addition, we propose
a set of techniques that we collectively call hierarchical compression for reducing the
space consumption of the hub labels. Visibility subdivisions show promising experimental results but additional optimizations are required to make them competitive with the
highest-resolution grid employed by Du et al. (2023). Likewise, the employment of hierarchical compression in the same experimental setting exhibits a remarkable reduction
in the space consumption of hub labels in visibility subdivisions but incurs a penalty
in the form of a minor increase in query times. Finally, we present the results of our
work prior to the revision of our objective induced by the aforementioned work by Du
et al. (2023). This includes a revised theoretical description of the ESPP algorithm
Polyanya(Cui et al., 2017) and the proposal of an algorithm called Neighborhood Search
which, using preprocessing, reports for a query point the set of obstacle vertices visible
from that point. Experimental results suggest that in its current form, Neighborhood
Search is inferior to competing algorithms.

Supervisor(s)  

Jacob Holm

External examiner(s)  

Eva Rotenberg

Date and time  

13-06-2023 from 13:00

Room  

A105 at HCØ, Universitetsparken 5

Name of student(s)  

Andreas Bjerregaard Jeppesen

Study Programme  

Computer Science

Title  

Save the mice: in-silico perturbation of genes in deep generative models

Abstract  

Abstract: Generative decoders provide more than just an output. By leveraging
their gradients with respect to the latent dimensions, we enable a
new toolbox which can: i) determine gene-gene correlations from a
single sample, ii) infer arbitrary perturbation effects on interpretive
flow maps, and iii) establish a similarity metric based on latent space
geodesics which parallel pseudotime. The developed methods demonstrated an ability to identify co-regulated gene groups and infer the
effects of gene knockdowns, overexpression, and cardiotoxin response
in mice. Furthermore, the application of latent space geodesics outperform simple linear methods in estimating pseudotime in C. elegans
embryogenesis. This thesis reveals new potentials for old models in
the complex world of single-cell biology.

Supervisor(s)  

Anders Krogh

External examiner(s)  

Jes Frellsen

Date and time  

13-06-2023 at 14:00

Room  

Seminar room, Panum 33.4.D

Name of student(s)  

Nicole Kozlová

Study Programme  

Computer Science

Title  

Multi-modal brain MRI segmentation using deep learning-based registration and transfer learning

Abstract  

Abstract. Magnetic resonance imaging (MRI) is, due to its high spatial
resolution, an important scan technique used in neurology and neurosurgery to diagnose and treat various brain disorders. One of the challenges of MRI is linking information from different images to aid in diagnosis and treatment. In this paper, we present a study on performing anatomical segmentation on Fluid Attenuated Inversion Recovery
(FLAIR) images, which are the most commonly used scans for searching for abnormalities. To perform the segmentation, we use a 2.5D-based
deep learning method proposed for fast, automatic, and accurate segmentation of the T1-weighted scan of the human brain into 129 regions. Then
we performed rigid registration of T1 and FLAIR images using a convolutional neural network (CNN), learning the transformation parameters.
Our study demonstrates the potential of deep learning-based methods
for accurate anatomical segmentation and registration of multi-modal
MRI images, leading to improve diagnosis and treatment of various neurological disorders.
Keywords: Deep learning · FLAIR · T1 · Segmentation · Registration.

Supervisor(s)  

Mads Nielsen, Mostafa Mehdipour Ghazi

External examiner(s)  

Rasmus Reinhold Paulsen

Date and time  

14/6 - 9:00-10:00

Room  

ØV3

Name of student(s)  

Kasper Unn Weihe

Study Programme  

Computer Science

Title  

Convex Optimization and Parallel Computing for Portfolio Optimization

Abstract  

Abstract
This thesis proposes a parallel portfolio optimization strategy that integrates environmental, social, and corporate governance (ESG) factors, while also balancing risk and return. Convex optimization techniques, particularly for constrained
problems, form the basis for addressing this challenge.
To efficiently compute multi-dimensional frontiers, symbolizing optimal portfolios, the high-performance parallel programming language, Futhark [1], is utilized. The research addresses the high computational demands inherent in solving
either a large grid of convex optimization problems or a single, very large problem,
harnessing general-purpose graphics processing units (GPGPUs). Different parallel computing strategies are investigated to enhance computational efficiency.
The investigation seeks to develop a data parallel method to solve convex optimization problems, with a focus on portfolio optimization as a case study. The
research aims to investigate how convex optimization can be parallel and applied
to portfolio management.
The findings of this research are anticipated to provide valuable contributions to
the domains of finance and computer science, presenting insights into the application of convex optimization techniques and parallel computing for investment
strategies. The relevance of solving large multi-dimensional optimization problems extends beyond portfolio optimization and can be applied to a multitude
of problems in diverse fields. Additionally, this thesis examines the efficacy of
Futhark as a high-performance parallel programming language for solving convex
optimization problems. Three Futhark modules are presented; the first one is for
solving linear systems of equations, which is used as a building block for the second module, a convex optimization module, capable of solving multi-dimensional
optimization problems, while the third is a portfolio optimization module, tailored
to optimize portfolios based on various constraints, which is implemented using
the aforementioned convex optimization module.
Keywords: Convex optimization, Portfolio Optimization, Futhark

Supervisor(s)  

Martin Elsman

External examiner(s)  

Mads Rosendahl

Date and time  

14-06-2023 at 9:00-10:00

Room  

SCI-DIKU-UP1-2-0-04

Name of student(s)  

Yihe Zhang

Study Programme  

Computer Science

Title  

Cross-Vendor Analysis of Mammographic Texture Model: A Longitudinal Study on Breast Cancer Risk Prediction

Abstract  

Abstract
This study presents an innovative deep learning model for predicting breast
cancer risk utilizing longitudinal data from two rounds of mammographic
screenings. The primary aim was to evaluate the impact of longitudinal data
on the model’s performance. Our findings showed high consistency in risk
scores across different mammogram views, laterality, and over time, suggesting the model’s stability and reliability. This consistency held true for both
healthy women and those diagnosed with various cancer types. Additionally,
we observed that the model’s risk scores were not significantly affected by
mammogram views (MLO or CC) or laterality (left or right breast), further
attesting to its robustness. Despite minor temporal variations in risk scores
across screening rounds and age groups, these did not significantly alter the
overall risk assessment, indicating that the model’s risk assessment remains
stable over time. The model’s stability, reliability, and resistance to various
factors make it a promising tool for improving breast cancer risk prediction and contributing to more personalized and efficient screening strategies.
However, the potential impact of sample size, population characteristics, and
the model’s robustness when applied to combined datasets from multiple
screening rounds warrant further exploration.

Supervisor(s)  

Mads Nielsen, Andreas Lauritzen

External examiner(s)  

Rasmus Reinhold Paulsen

Date and time  

14-06-2023 at 10:00-11:00

Room  

ØV3

Name of student(s)  

Yu Wang

Study Programme  

Computer Science

Title  

The evaluation of Multi-CLIP model on IGLUE

Abstract  

Abstract
Multilingual and multi-modal researches in the Natural Language Processing area are quite popular nowadays, such as building multilingual and multi-modal models, creating datasets, proposing
relevant benchmark tasks, and so on. The CLIP ViT-B/32 XLM-Roberta based model (henceforth, Multi-CLIP model) is pre-trained on the LAION-5B dataset, a large-scale dataset containing
5B multilingual text-image pairs. Therefore, the Multi-CLIP model can be used in several multilingual and multi-modal tasks and performs very well. The IGLUE benchmark tasks contains four
different multilingual and multi-modal tasks (visual natural language inference, visual question
answering, visual reasoning, and cross-modal retrieval) to evaluate models from different aspects.
This project aims to evaluate the performance of the Multi-CLIP model on these tasks, and this
report shows the Multi-CLIP model achieves promising results on visual natural language inference and cross-modal retrieval tasks, and achieves bad results on other tasks. I provide the code
of the project at https://github.com/yu-tracy/Multi-CLIP-model-on-IGLUE-benchmark

Supervisor(s)  

Desmond Elliott

External examiner(s)  

Johan Kjeldgaard-Pedersen

Date and time  

14-06-2023 at 11:00-12:00

Room  

Meeting room, 4th Floor of Vibenshuset

Name of student(s)  

Magnus Diamant and William Lundsgaard Pedersen

Study Programme  

Computer Science

Title  

Automatic Quality Assessment and Pathology Detection of Magnetic Resonance Imaging

Abstract  

Abstract
Brain segmentation, crucial for analyzing medical images, can be slow and expensive
when performed manually. On the other hand, automated methods offer speed, but
their performance can suffer in the presence of low-quality scans or brain pathologies.
This study, conducted on 35,818 3D brain Magnetic Resonance Images (MRIs) from
24,261 patients in the capital region of Denmark, primarily focuses on detecting such
significant pathologies and the quality assessment of the associated segmented images.
Both pathology detection and segmentation failure prediction are facilitated by a unified pipeline, where 3D MRIs are segmented into 132 regions using a U-net-based model,
generating a vector representation for classifier input. Concurrently, processed and segmented MRIs are input into a pre-trained Residual Neural Network (ResNet), serving
either as a feature extractor with a classifier or, in a unique procedure, as a fine-tuned
combination with a Neural Net.
In pathology detection, we achieved an area under the receiver operating characteristic
curve (ROC-AUC) score of 0.85, indicating that Support Vector Machine (SVM) was
the most effective classifier for volumetric brain data. Furthermore, a finetuned ResNet
combined with a Neural Net achieved a ROC-AUC score of 0.99 in predicting segmentation failure on segmented brain MRIs, mirroring the performance of SVM on BV for
the same task. These findings suggest that adding complexity by integrating processed
MRIs and utilizing a pre-trained ResNet does not enhance results but notably escalates
computational time.
This study emphasizes the importance of model selection and further refinement for
effective medical image analysis and disease detection.

Supervisor(s)  

Mads Nielsen and Kiril Vadimovic Klein

External examiner(s)  

Rasmus Reinhold Paulsen

Date and time  

14-06-2023 at 13:00-14:30

Room  

ØV3

Name of student(s)  

Asbjørn Munk

Study Programme  

Computer Science

Title  

Domain Adaptation of U-Nets With Applications in Hippocampus Segmentation

Abstract  

Abstract
The current state-of-the art techniques for medical image segmentation are often based on
U-Net architectures. Despite the powerful performance, the architecture often does not perform well when used on data which has different characteristics than the data it was trained
on. One promising solution to tackle this problem is unsupervised domain adaptation. While
many unsupervised domain adaptation methodologies in the medical domain have been developed, the work is only loosely based on theory. This thesis concerns itself with domain
adaptation for medical image segmentation from both theoretical and practical perspectives.
This thesis outlays the theoretical foundation, from which practical methodologies can be
build. Next, we propose a theoretically grounded unsupervised domain adaptation framework for U-Nets based on the Margin Disparity Discrepancy called the MDD-UNet. The
MDD-UNet is able to learn features which are domain invariant with no knowledge about
the labels in the target domain. We evaluate the proposed technique on the task of hippocampus segmentation from brain MRI images, and find that the MDD-UNet outperforms
both the regular U-Net and the U-Net when trained with augmentations.

Supervisor(s)  

Mads Nielsen

External examiner(s)  

Rasmus Reinhold Paulsen

Date and time  

14-06-2023 from 14:30-15:30

Room  

ØV3

Name of student(s)  

Thomas Schrum Nicolet and Karl Emil Levinsen

Study Programme  

Computer Science

Title  

Domain-specific Diffusion in the Butterfly Domain

Abstract  

Abstract
Diffusion-based generative models have made breakthroughs in generating synthetic image
data, ranging from photo-realism to images of artistic merit. Large-scale state-of-the-art models
work in the general domain, however, and being able to generate synthetic compositions of
virtually anything potentially comes at the cost of specificity. In this thesis, we explore how
to make a diffusion model specific to the butterfly domain. In doing so, we aim to explore the
theory and implementation of diffusion models in general, as well as ways of evaluating such
models. To do this, we finetune a pre-trained state-of-the-art model as well as train a model
purely on our own data. We include both quantifiable metrics as well as a qualitative human
evaluation. We conclude that our model finetuned on a pre-trained checkpoint performs best,
and suggest an efficient way to develop domain-specific text-to-image diffusion models.

Supervisor(s)  

Stefan Sommer

External examiner(s)  

Dan Witzner Hansen

Date and time  

15-06-2023 at 9:00

Room  

Seminar room Østervoldgade 3

Name of student(s)  

Binbin Huang and Ziqian Li

Study Programme  

Computer Science

Title  

Development of a tool for extracting event logs from natural language case descriptions

Abstract  

Abstract
The utilization of natural language event extraction models has increased significantly
in recent years. To structure the model processing, pipelines are a widely adopted approach. While the renowned Scikit-learn library offers an implementation of pipelines,
it still has some limitations. In this study, we address these limitations by extending
the functionality of the Scikit-learn pipeline library. Our extensions introduce a set
of useful APIs empowering users to define more general and complex pipelines and
access intermediate results. Furthermore, we develop a user-friendly graphical user
interface (GUI) tool that simplifies the pipeline definition process in an intuitive way
based on the extended pipeline APIs The GUI tool facilitates real-time visualization
of the pipeline structure and its associated intermediate results, enhancing the overall
usability and effectiveness of event extraction model development.
Keywords: Pipeline, Natural Language Processing, Event Extraction

Supervisor(s)  

Tijs Slaats

External examiner(s)  

Søren Debois

Date and time  

15-06-2023 at 09:00-10:30

Room  

SCI-DIKU-sigurdsgade-2-03

Name of student(s)  

Casper Stybert and Jasper Neo Lassen

Study Programme  

Computer Science

Title  

Implementing a Blockchain with Native Support for DCR Graph Smart Contracts

Abstract  

Abstract
This master’s thesis proposes a novel blockchain environment which
integrates support for DCR Graphs (Dynamic Condition Response Graphs)
natively. Previous research has explored integrating DCR Graphs with
Solidity in Ethereum, but encountered performance inefficiencies. Our
solution improves performance and scalability of such an integration by
reducing the amount of overhead for DCR Graph specific operations,
while offering extendibility and optimization potential. This paper lays
out the background theory of the blockchain technology and the DCR
Graph process modeling language, which is necessary to understand for
implementing a working solution. Based on this foundation of knowledge,
the design of the solution is then presented, followed by the implementation of this design in C#. Using this implementation, performance
evaluations demonstrate significant improvements for transaction validation when compared to previous results using Ethereum as a blockchain,
while the impact of certain implemented technical features are analyzed
and showcased. Following the results, the performance, extendibility,
applicability, and potential improvements of the solution are discussed.
This all leads into the conclusion that the proposed solution in this paper
is a novel, extensible contribution to the field of research which integrates
DCR Graphs into a custom blockchain environment, with clear performance improvements when compared to results in previous research.
While the solution as a whole needs to see improvements to be usable
for commercial use, it serves as a stepping stone for future research.

Supervisor(s)  

Tijs Slaats

External examiner(s)  

Søren Debois

Date and time  

15-06-2023 from 11:00-13:00

Room  

SCI-DIKU-sigurdsgade-2-03

Name of student(s)  

Michaela Isabella André Edwards

Study Programme  

Computer Science

Title  

Representing and understanding DCR graphs in 3D

Abstract  

Abstract
Processes with a high degree of variability can be difficult to understand and
work with for the people who are managing them. DCR graph notation can
be used to create concise graphical models of these types of processes. There
are digital tools available for modeling DCR graphs in 2D, however, little to
no research has gone into 3D DCR graphs. Thus, we explore whether the
addition of a third dimension may provide value to end-users. In this project,
we seek to answer the question of whether we can support users in modeling
and understanding processes by providing an immersive virtual reality (VR)
environment for creating 3D DCR graphs. We describe the development process of a prototype for a 3D DCR graph modeling tool, emphasizing our main
challenges and solutions for them. We also implement VR support and validate the prototype through experimentation and semi-structured interviews
with three domain experts from the industry and academia. We found that
users showed an interest in the potential of the tool, especially for collaboration, and they provided multiple ideas for expanding the tool so it becomes
more practical and useful.
Keywords: DCR graphs, 3D declarative processes, Virtual Reality, 3D process models

Supervisor(s)  

Tijs Slaats

External examiner(s)  

Søren Debois

Date and time  

15-06-2023 at 13:15-14:15

Room  

SCI-DIKU-sigurdsgade-2-03

Name of student(s)  

Markus Krabbe Larsen

Study Programme  

Computer Science

Title  

Adjoint Affine Interpretation

Abstract  

Abstract
We define a symbolic language of functions on combinatory form with a
sublanguage of linear functions as a specific instantiation of the framework proposed by Elsman et al. in “Combinatory Adjoints and Differentiation”. With
this language we can define an affine interpreter that returns symbolic linear functions, which can then be turned around by calculating their adjoint
symbolically. The interpreter is proven correct with regard to its denotational
semantics, and we show how to apply the adjoint affine interpreter to train a
neural network.

Supervisor(s)  

Fritz Henglein

External examiner(s)  

Mads Rosendahl

Date and time  

16-06-2023 at 10:00-11:00

Room  

PLTC meeting room (772-01-0-S29)

Name of student(s)  

Trine Dag Randløv

Study Programme  

Computer Science

Title  

Toward a Monadic Functional Machine Model for Computability and Complexity Theory: Finite and Pushdown Automata

Abstract  

Abstract
This thesis explores the use of monadic functional machine models to bridge the gap between functional programming and computational models, focusing on Finite Automata (FAs) and Pushdown Automata (PDAs). By integrating monadic functional programming, we aim to provide
more expressive and modular representations of FAs and PDAs. The thesis develops functional
machine models for FAs and PDAs and demonstrates equivalence between deterministic and
nondeterministic FAs. Additionally, we propose a purely functional alternative, the two-stack
pushdown automaton (2sPDA). Future work involves expanding functional implementations to
include nondeterministic- and monadic generalized pushdown automata and deterministic/nondeterministic Turing machines. This thesis establishes a foundation for constructing and analyzing
machines within computational models, enhancing our understanding of their capabilities. It also
provides tools for teaching about said computational models.

Supervisor(s)  

Fritz Henglein and Robert Schenck

External examiner(s)  

Mads Rosendahl

Date and time  

Friday, June 16th at 11-12

Room  

PLTC meeting room (772-01-0-S29)

Name of student(s)  

Adam Zdziechowski and Klaus Philipp Theyssen

Study Programme  

Computer Science

Title  

Monitoring Safety Properties in Event-Driven Microservices

Abstract  

Abstract
Large parts of our modern society depend on software, therefore creating dependable software artifacts is an important problem.
At the time of writing microservice architectures are a widespread approach
for constructing software systems. These architectures promise several benefits
like scalability, fault tolerance, improved resource utilization in the cloud and
an improved development process across multiple engineering teams. At the
same time implementing the microservices architecture comes with several
challenges and potential pitfalls that naturally arise in distributed systems.
In this thesis, we are concerned with a certain class of microservices, namely
event-driven architectures, in which the individual components communicate
via asynchronous events using a messaging system.
An inherent challenge in these systems is dealing with data integrity, as each
service typically uses its own data store. Due to a lack of a centralized component that enforces constraints, one might experience faulty replication and
inconsistencies in the data.
To address this issue we created a simulation tool for microservice applications
with the goal of quantifying data integrity violations. For this, we have created
an invariant definition language and a workload generator to execute specific
workloads against the system under test. The envisioned use by a developer
is to define key invariants about an application and a workload scenario. The
tool translates the invariants into stream processing operators which monitor
the invariants while executing the workload. The resulting metrics are visualized and presented to the user in a dashboard. The tool aims to allow fast
experimental investigations of the data integrity properties of the application.
This in turn lets the developer make informed decisions about how to evolve
an application architecture, to ensure costly anomalies are discovered before
the application is deployed into the production environment.

Supervisor(s)  

Yongluan Zhou

External examiner(s)  

Philippe Bonnet

Date and time  

20-06-2023 at 10:00

Room  

Sigurdsgade 41, 2.03

Name of student(s)  

GuanRan Tai

Study Programme  

Computer Science

Title  

An open source library for atomic migration of Erlang processes

Abstract  

Abstract
Resource management in computer systems relies on processes to control program behavior.
Some programming languages use this concept to provide developers with efficient and convenient coding procedures. In the Erlang VM, the runtime environment for the Erlang and Elixir
programming languages, developers can use lightweight processes called Erlang processes.
However, the current Erlang VM lacks a solution for seamless, atomic migration of an Erlang
process while preserving the requirement order property. In this paper, we propose a novel
mechanism for migrating Erlang processes and present the design and implementation of an
Elixir library based on this mechanism. In addition, seven group performance tests are conducted to analyze the behavior of the library. The tests examine different scenarios, such as
increasing the number of clients or requests, to evaluate the performance of the library during migration. A comparison is made between behavior with and without the migration library,
which provides insight into its effectiveness. And by integrating the library into a real product
for testing, which proves its correctness.

Supervisor(s)  

Yongluan Zhou

External examiner(s)  

Philippe Bonnet

Date and time  

20-06-2023 at 11:00

Room  

Sigurdsgade 41, 2.03

Name of student(s)  

Bo Cui

Study Programme  

Computer Science

Title  

Failure Recovery in Deterministic Transaction Processing Systems

Abstract  

Abstract
This thesis presents a novel approach for enhancing failure recovery in actor-based
deterministic transaction processing systems. As the reliability and robustness of
these systems are critical in today’s digitalized world, the development of efficient
and adaptable recovery mechanisms is of utmost importance. To address this, we
have devised a specialized recovery actor and a comprehensive recovery process,
implementing five unique recovery modes to handle various failure scenarios.
Each mode is extensively analyzed, and its performance is empirically evaluated,
providing insights into their respective strengths and potential applications. Our
results indicate improvements in system resilience and adaptability, contributing to
minimized system downtime and maximized data integrity

Supervisor(s)  

Yongluan Zhou and Yijian liu

External examiner(s)  

Philippe Bonnet

Date and time  

20-06-2023 at 13:00

Room  

Sigurdsgade 41, 2.03

Name of student(s)  

Jacob Christian Herbst

Study Programme  

Computer Science

Title  

Verifying eBPF programs in the Linux Kernel

Abstract  

Abstract
eBPF is a Linux kernel sub-system that allows users to run code of a limited assemblylike language inside the Linux kernel. This can potentially be dangerous in the hands of
malicious or even uncareful users. To combat this, the Linux kernel provides static analysis
of eBPF programs to reject harmful programs. This static analysis has in the past shown to
be unsound, meaning harmful programs have been accepted. In this project, we investigate
the feasibility of making a logical proof checker that runs inside the kernel, implemented
in Rust. We do so by considering the Logical Framework with Side Conditions (LFSC)
language. The motivation behind such a proof checker is to be able to formally prove the
correctness of eBPF programs. We consider the feasibility not only as a standalone tool
but as a hypothetical part of a larger Proof Carrying Code (PCC) architecture. By this, we
provide an analysis of how eBPF programs are loaded and what the static analysis consists
of and hereby describe a vision for a PCC architecture. The main focus of the report is to
describe and evaluate the implementation of LFSC with respect to eBPF, PCC, and Rust in
the Linux kernel. Although the implementation is not as fast as other LFSC proof checkers
it does show decent promise for a part of a PCC system for eBPF.

Supervisor(s)  

Ken Friis Larsen

External examiner(s)  

Carsten Elmar Schürmann

Date and time  

20-06-2023 at 15:00

Room  

Room 01-0-029 at HCØ

Name of student(s)  

Leeann Quynh Do

Study Programme  

Computer Science

Title  

Image Segmentation of Sub-cellular Structures

Abstract  

Abstract
In the field of medical image analysis, it is very desirable to be able to
segment the cell walls in images of sub-cellular structures, to further apply
useful analyses on a single cell.
Images of sub-cellular structures has been provided, and methods for
segmenting the cell walls are proposed and investigated, including existing
convolutional neural networks, and a locally orderless network, which will
be proposed in combination with a convolutional neural network.
It is shown, that a convolutional neural network with a U-Net architecture using a ConvNeXt structure as a backbone, outperforms the proposed
method and other existing methods, with a F1 score of 0.9959.

Supervisor(s)  

Jon Sporring

External examiner(s)  

Rasmus Reinhold Paulsen

Date and time  

21-06-2023

Room  

Universitetsparken 1

Name of student(s)  

Lukas Mikelionis

Study Programme  

Computer Science

Title  

Structured Argument Mining in Persuasive Online Articles on Sustainable Diets

Abstract  

Abstract
Effectively encouraging behavioural change in terms of dietary preferences is
a challenging endeavour. It requires a good understanding of the society and
depends on convincing communication as well as sure actions from those in
power.
Argument Mining (AM), a sub-field of Natural Language Processing (NLP),
can be utilized to extract relevant information from texts, thus, for example,
helping to address some of the misconceptions surrounding the plant-based diets
and provide compelling arguments for switching toward more sustainable food
options, benefiting both, people and the environment.
This thesis introduces a new structured argument mining corpus on the topic
of sustainable diets. The dataset stands out from the existing corpora due to
previously unexplored content type of persuasive online articles. Additionally,
the corpus was annotated using crowdsourcing, illustrating that it is feasible to
gather reasonable quality data with this approach on a cognitively taxing task
like argument mining. However, compromises had to be made with respect to
the diversity and the representation of the participants.
The recent state-of-the-art end-to-end structured argument mining model is
examined on the corpus. Several training configurations were employed, including training solely on the target corpus and, alternatively, pre-training on
auxiliary AM corpora and subsequently fine-tuning on the target dataset. The
phenomenon of cross-corpora knowledge transfer was observed, as the model
trained with the latter method achieved higher performance than the former,
provided that the auxiliary and target corpora are sufficiently similar.
iv
Lastly, while the achieved scores were not impressive, at least when compared
to what is reported in the literature on other AM corpora, the quantitative
analysis does not tell the full story. The lower performance appears to be related
to the complexity of the task on the content type of persuasive online articles.
The complexity itself likely stems from the subjectiveness of the task, which is
supported by a relatively low inter-annotator agreement.

Supervisor(s)  

Daniel Hershcovich

External examiner(s)  

Christian Hardmeier

Date and time  

22-06-2023 at 9:00

Room  

NLP meeting room, 4th floor, Vibenshuset

Name of student(s)  

Mengqian Wang

Study Programme  

Computer Science

Title  

Implicit knowledge in CV/NLP models

Abstract  

With the large pre-trained language models making huge progress in various downstream NLP tasks, the investigation about if
pre-trained language models could be used as knowledge bases is going on. Previous studies also showed that commonsense knowledge
could be extracted from pre-trained vision models[34]. In this work, I would use linear mapping to relate the language or visual
representations of knowledge concepts with the reference knowledge graph embedding space. The experiment demonstrates that the
alignment precision will increase with the pre-trained model size and also be affected by concept category and image dispersion.
CCS Concepts: • Computing methodologies → Knowledge representation and reasoning; Natural language processing;
Computer vision.
Additional Key Words and Phrases: knowledge base, linear mapping

Supervisor(s)  

Anders Søgaard

External examiner(s)  

Zeljko Agic

Date and time  

22-06-2023 at 10:15

Room  

Online

Name of student(s)  

Mathias Lykke Gammelgaard and Jonathan Gabel Christiansen

Study Programme  

Computer Science

Title  

Large language models converge toward human-like concept organization

Abstract  

Abstract
Large language models show human-like performance in knowledge extraction, reasoning and dialogue, but it remains controversial whether this performance simply
reflects memorization and pattern matching, or whether it reflects understanding,
i.e., modeling of the environment and reasoning with model representations. We
show that large language models learn to organize concepts in ways that are strikingly similar to how concepts are organized in knowledge bases. Knowledge bases
model collective, institutional knowledge, and large language models seem to be
able to induce this knowledge from raw text. We present convergence results across
four families of language models and three knowledge graph embeddings.

Supervisor(s)  

Anders Søgaard

External examiner(s)  

Zeljko Agic

Date and time  

22-06-2023 at 9:30

Room  

Online

Name of student(s)  

Fengmao Wang

Study Programme  

Computer Science

Title  

Hierarchical Graph Transformers for Legal Judgment Prediction

Abstract  

Abstract
In the legal domain, texts often exceed the
length that can be efficiently processed by
general Large Language Models. To address
this challenge, non-hierarchical sparse attention
Transformer-based models, such as Longformers, have gained popularity for handling long
documents. Despite their popularity, these models often suffer from a lack of computational
efficiency and require substantial GPU memory
resources. Based on such observation, I propose a new model adapted from Hierarchical
Attention Transformers. It incorporates Graph
Transformer as the segment-wise encoders, and
BERT Layer as the cross-segment encoders.
Documents are processed in a graph-based
manner, in which each document is parsed
into Abstract Meaning Representation graphs.
Based on the experiment results and analysis,
my approach shows potential comparable to
Longformers.

Supervisor(s)  

Daniel Hershcovich, Ilias Chalkidis

External examiner(s)  

Christian Hardmeier

Date and time  

22-06-2023 at 10:00

Room  

NLP meeting room, 4th floor, Vibenshuset

Name of student(s)  

Jiaang Li

Study Programme  

Computer Science

Title  

Large Language Models Converge on Brain-Inspired Representations

Abstract  

Abstract
Currently humans interact with computers through a graphical interface or, in some cases,
through verbal commands. We can type or speak commands and queries, and then the computer can respond with text or speech. But how does the computer analyze and generate
language? One of the ways that researchers tackle this challenge is by using large language
models (LMs), which are trained on large amounts of text to predict masked or forthcoming
words. And these models have recently been shown to generate activations similar to those of
the human brain. Decoding language stimuli from brain recordings aims to deepen our understanding of the human language system and build a solid foundation for bridging human and
natural language processing through the Brain-Computer Interface. Here, we delve into the
alignment between language models and the human brain and present how language models
acquire similar representations of concepts as the human brain does.

Supervisor(s)  

Anders Søgaard

External examiner(s)  

Zeljko Agic

Date and time  

22-06-2023 at 12:00

Room  

Online

Name of student(s)  

Sebastian Hammer Eliassen

Study Programme  

Computer Science

Title  

Activation Compression on Graph Neural Networks

Abstract  

Abstract
This project tackles the challenge of increasing dataset and parameter sizes in
machine learning, particularly within Graph Neural Networks (GNNs). The
project first looks at an existing method for compressing the memory usage of
GNNs, which reduces the memory footprint by 32× while only having a 0.2% −
0.5%− accuracy drop. We replicate the results of this paper, and extend them
further by including a new performance metric: energy expenditure. Here we see
that even when slight compression is used, there is always a 2-3 times increase in
energy, compared to no compression at all, warranting a further look into how
this overhead can be reduced. This study also introduces grouped quantization
for GNNs, further reducing memory usage by 16% and improving training speed
by around 5%. To reduce accuracy loss, we propose an approximation of the
distribution of activations, achieving an r-value of 0.99981, significantly higher
than the current approximation, which has an r-value of 0.99094. Using this
approximation we attempt to minimize variance induced by the compression
method. This results in negligible improvements, suggesting further research in
this area is needed.

Supervisor(s)  

Raghavendra Selvan

External examiner(s)  

Lee Herluf Lund Lassen

Date and time  

22-06-2023 at 13:00

Room  

SCI-DIKU-UP1-2-0-15 (UP1).

Name of student(s)  

Frederik Alexander Noe and Jonas Friis

Study Programme  

Computer Science

Title  

Powerline detection in 3D point clouds from airborne LiDAR

Abstract  

Abstract
The Danish Agency for Data Supply and Infrastructure needs a system to effectively
monitor the powerlines to ensure reliable operation and prevent potential hazards. We
develop a highly extendable and easy to use end-to-end framework for segmenting
powerlines in 3D Point Clouds from airborne LiDAR. Real world applications of airborne
LiDAR create big-data. We employ the Minkowski Engine, a library for sparse tensors
that enables efficient deep learning on sparse 3D data in contrast to using a classic
3D Convolutional Neural Network for processing dense 3D data. We develop two
methodologies for relevant data selection. We reduce the dataset on average by 69.9%
and 95.4%, using the rule-based method and multi-view method, respectively. The
rule-based method is locked on a predefined rule-set from domain knowledge, however
the multi-view method is adjustable to extract wider range of relevant data points.
The experiment for the final 3D Sparse CNN is exploratory and not decisive. The
3D segmentation model obtains a mIoU score of 0.963 and mAcc score of 0.970. A
significant improvement from the existing baseline scoring 0.888 and 0.891, respectively.
For all related code, visit our Github repository: Link to repository.

Supervisor(s)  

Christian Igel

External examiner(s)  

Lee Herluf Lund Lassen

Date and time  

22-06-2023 at 15:30

Room  

SCI-DIKU-UP1-2-0-15

Name of student(s)  

Thor Hannibal Valsgaard

Study Programme  

Computer Science

Title  

Using Behavioural Cloning to Slay the Spire

Abstract  

Abstract
Slay the Spire is a turn-based video game. There are a lot of videos on YouTube of competent
Slay the Spire players. In this work I first present the game and the existing programs that
can play the game without human input. I then present my program, which can analyse
videos of people playing Slay the Spire and find out which actions were taken in which game
states. I then use that program on a set of videos from one of the best players and train
a model on the data derived from those videos. That model is then used to play the game
independently. The model was able to beat the game, but it was not better than any of the
other similar programs. I then tried to use reinforcement learning to train a similar model.
That model clearly performed better than random guessing, but it was still worse than the
first one.

Supervisor(s)  

Silas Nyboe Ørting

External examiner(s)  

Veronika Vladimirovna Cheplygina

Date and time  

23-06-2023 at 13:00

Room  

SCI-DIKU-UP1-2-0-06

Name of student(s)  

André Oskar Andersen

Study Programme  

Computer Science

Title  

Temporal Smoothing in 2D Human Pose Estimation for Bouldering

Abstract  

Abstract
In this thesis we implement four architectures for extending an already developed keypoint
detector for bouldering. The three architectures consist of (1) a single 3-dimensional convolutional layer followed by the ReLU activation function, (2) DeciWatch by Zeng Et al. [31], and
(3) two kinds of bidirectional convolutional LSTMs inspired by Unipise-LSTM by Artacho and
Savakis [3], where the difference between the two architectures lies in how they combine the
two processing directions. The models are pretrained on the BRACE dataset [19] and parts of
the Penn Action dataset [33], and further finetuned on a dataset for bouldering. The keypoint
detector and the finetuning dataset are both provided by ClimbAlong at NorthTech ApS. We
perform various experiments to find the optimal setting of the four models. Finally, we conclude, that DeciWatch by Zeng Et al. [31] yields the most accurate results, one of the bidirectional convolutional LSTMs yields the best rough estimations, and the simple 3-dimensional
convolutional layer yields the best results when also considering in the size and prediction
time of the models.

Supervisor(s)  

Kim Steenstrup Pedersen

External examiner(s)  

Rasmus Reinhold Paulsen

Date and time  

26-06-2023 at 09:00-09:55

Room  

UP1, Room 2-0-06

Name of student(s)  

Linas Einikis

Study Programme  

Computer Science

Title  

Herbarium sheet label data digitisation using handwritten text recognition

Abstract  

Abstract
Despite ongoing research efforts, the problem of recognizing offline handwritten historical text documents remains unsolved. There is a significant
demand for a functional solution, as it would greatly benefit cultural heritage
archives and museums. This work aims to design and implement the architectural components of an end-to-end transcription pipeline using limited
GPU resources.
The implemented line segmentation system is based on a baseline localisation module utilizing a state-of-the-art residual U-Net architecture with a
redesigned post-processing step. The proposed changes established a more
efficient processing procedure with fewer execution steps while still maintaining comparable precision.
For the transcription module, considering the large architecture size and
extensive training set requirements for the current state-of-the-art methods,
alternative approaches were explored. Initially, it was hypothesized that
a hybrid CNN and Transformer encoder could potentially outperform the
Vision Transformer variant in capturing visual text features with a smaller
training dataset. However, the comparative study revealed that both encoder
variants achieved similar results. The hybrid architecture with 11M and 38M
parameter settings yielded a Character Error Rate (CER) score of 37.22%
and 34.05% respectively. On the other hand, the DeiT Vision transformer
resulted in a CER of 33.85%.
The overall best transcription performance for the Natural History Museum of Denmark (NHMD) historical data was obtained by a warm-started
and fine-tuned TrOCRBASE method resulting in 11.93% CER and thus highlighting the style variation resilience of architecture which is achievable with
a relatively small fine-tuning procedure.
The visual inspection highlighted that the proposed end-to-end pipeline
achieves human-like transcription performance, however, it still contains errors. The predictions for some samples suggest that a large language model
decoder could not fully learn the specific Greenland herbarium context with
a limited number of training samples. Besides, there were errors made due to
the questionable ground truth entries. To further improve the results, there
is a need for a strict prediction error handling protocol. Otherwise, the score
could be advanced by building on top of synthetic data generation ideas and
making a larger training dataset with representative historical-style text.

Supervisor(s)  

Kim Steenstrup Pedersen

External examiner(s)  

Rasmus Reinhold Paulsen

Date and time  

June 26 at 10:00 - 10:55

Room  

DIKU Universitetsparken 1, Room 2-0-06.

Name of student(s)  

Ben Sauber

Study Programme  

Computer Science

Title  

Comparing explainability methods for session-based recommender systems

Abstract  

Abstract
Recommender systems allow to quickly produce highly personalized content.
The benefits of these recommendations depend on the domain. For e-commerce,
personalized content can lead to purchases. For other domains such as finance
or law, the main benefit might be automating an otherwise time costly process.
A central idea regarding recommender systems was that they work best
given a lot of historical data about the users and items. Session-based recommenders on the other hand are trained on short-term user interactions and
do not rely on long-term user profiles. With the rise of privacy regulations,
session-based systems become more important as they only need information about the ongoing session [54], as well as previous sessions which can be
anonymized. This means that the recommender does not connect the user of
the current session to their previous sessions when making recommendations.
Explainability has become an important point of discussion as AI keeps
getting integrated into the lives of billions of people. However, in most cases
recommender systems are still treated like black-boxes and do not offer explanations to the end users. This is a growing concern as recommender systems
are used in crucial domains that affect people.
This thesis analyzes how current advances in explainable AI such as local interpretable model-agnostic explanations (LIME) and Shapley additive
explanations (SHAP) can be applied to the domain of session-based recommender systems. To do so, this thesis explores the relation between SHAP
and LIME by computing the similarity of the feature importance generated by
these two methods across different models and datasets. In addition, because
LIME and SHAP were not built for recommender systems, this thesis uses
a counterfactual approach to prove that there is a relation between the importance attributed to a feature and the top 20 recommendations when that
feature is modified.

Supervisor(s)  

Maria Maistro, Christina Lioma, Simone Borg Bruun

External examiner(s)  

Martin Aumüller

Date and time  

26-06-2023 at 13:00

Room  

Christina’s office: Universitetsparken 1, level 2, room number 1.2.26

Name of student(s)  

Peter Kanstrup Larsen

Study Programme  

Computer Science

Title  

Application of Probabilistic Machine Learning Methods for Protein Generation

Abstract  

Abstract
The recent introduction of denoising diffusion probabilistic models has led to a
surge in the field of AI. While text-to-image models such as OpenAI’s DALL-E
2 have recieved widespread public attention, many scientific fields are looking
towards the technology as well. Researches in the field of protein design recently
demonstrated that such models could be used for protein generation, achieving
results comparable to state-of-the-art models. In this thesis, we show how a
diffusion model can be implemented in Futhark, and compare our results to an
equivalent model implemented using the PyTorch library.

Supervisor(s)  

Cosmin Eugen Oancea

External examiner(s)  

Mircea Filip Lungu

Date and time  

27-06-2023 at 10:00-11:00

Room  

SCI-DIKU-HCO-01-0-029 (PLTC meeting room)

Name of student(s)  

Jingyu Huang

Study Programme  

Computer Science

Title  

Anatomical Prior-based Segmentation of Deep Brain Nuclei using Adversarial Training

Abstract  

Abstract
Accurate segmentation of brain regions is central to the investigation
of relevant signal in magnetic resonance imaging (MRI) of the brain. In
recent years, the emergence of susceptibility-based MRI techniques have been
shown to be particularly useful for investigating properties of the deep brain
nuclei which are affected in diseased states such as their iron or melanin
content. Owing to the lack of specialized tools, manual labelling remains the
most used segmentation approach for these nuclei. However, convolutional
neural networks have been recently introduced as an efficient and accurate
alternative to enable automatic segmentation. In cases presenting a low
contrast-to-noise ratio, convolutional neural networks have difficulty inferring
anatomically correct segmentations. To address this problem, we investigated
using anatomical priors through adversarial training as a simple and effective
strategy to improve segmentation accuracy in terms of shape and position.
This thesis is divided into two parts. The first part focuses on the
adversarial training of U-Net models. The experiments for U-Net revolve
around three experimental settings: adjusting the adversarial loss weight,
pre-training the generator, and binarizing the inputs to the discriminator.
This project focuses on the dentate nucleus as the most difficult region in
our test dataset due to its low contrast-to-noise ratio. By adjusting those
three experimental settings, the U-Net with adversarial training exhibited
an average improvement of 0.04 in Dice coefficient compared to the single
U-Net on the test set, while also achieving an average reduction of 1.8 mm
in Hausdorff distance.
To assess the generalizability and robustness of adversarial training
methods across different model architectures, we applied the experimental
settings that had yielded promising results on the U-Net also to the U-Net++
model. However, we found that employing the same parameters and experimental settings on U-Net++ resulted in poor segmentation performances.
Hence, in the second part of our study, we conducted experiments investigating the influence of hyperparameters and training strategies in order to
understand and provide a practical guidance for adversarial training.
In conclusion, this thesis presents a comprehensive investigation of
anatomical prior-based segmentation models using adversarial training. In
the first part, we improved the performance of U-Net on both Dice coefficient
and Hausdorff distance. In the second part, we provided a comprehensive
guideline for employing adversarial training from our experiments.

Supervisor(s)  

Primary supervisor: Melanie Ganz
Co-supervisor: Vincent Beliveau

External examiner(s)  

Oula Tapio Puonti

Date and time  

27-06-2023 at 16:00-17:00

Room  

SCI-DIKU-UP1-2-0-04

Name of student(s)  

Kristian Quirin Hansen

Study Programme  

Computer Science

Title  

Fine-Grained Bird Sketch Generation and Evaluation

Abstract  

Abstract
Recent advancements in deep learning have revolutionized the art world and our understanding of AI, particularly with recent work synthesizing highly detailed raster images through pixel
modification. In this thesis, I explore the generation of abstract vector sketches based on reference images through a stroke-based rendering (SBR) methodology on Beziér curves, leveraging
a differentiable vector graphics rasterizer for gradient-based optimization. Previous research
primarily focuses on generating coarse-grained object sketches. In contrast, this thesis focus
on generating fine-grained, simplified bird sketches using the zero-shot multimodal model CLIP
(Contrastive-Language-Image-Pretraining), where the aim is to strike a balance between abstraction and essential features in the rendering process to establish a sketch of a bird where its family
and species can be identified.
The thesis presents a zero-shot evaluation metric using CLIP to quantitatively assess the quality
of the rendered sketches by predicting the bird species and family of the bird in the synthesized
sketch. Various aspects of the rendering approach are examined to explore the potential of using
a stroke-based rendering approach for sketching fine-grained bird species. I demonstrate that initializing stroke points in relation to bird body parts in the reference image results in sketches with
a higher chance of capturing the bird’s essential features. Furthermore, I showcase the versatility of stroke-based rendering, from generating highly detailed to abstract synthesized sketches
through methods such as optimizing over color and width parameters of the Beziér curves in
the sketch to achieving more simplified and abstract sketches through regularization and postprocessing methods.

Supervisor(s)  

Serge Belongie and Stella Frank

External examiner(s)  

Jeppe Revall Frisvad

Date and time  

27-06-2023 at 10:30-11:30

Room  

Øster Voldgade 3, Mødelokale A

Name of student(s)  

Frederik Lunn Berthelsen and Kasper Erik Lindquist

Study Programme  

Computer Science

Title  

An Information Flow Analyser for eBPF

Abstract  

Abstract
eBPF is a technology allowing users to extend the Linux Kernels’ capabilities by injecting code
into the kernel, which runs when triggered by predefined hooks. This introduces various security
risks since it enables users to interact with kernel space. To mitigate these risks, eBPF introduces
a static analysis verifier for eBPF programs. This verifier focuses on the safetyness of programs
but not the confidentiality and integrity of programs. We introduce an information flow analysis
tool designed for eBPF. This tool enables users to verify that a program does not leak information
with respect to a given security policy. This tool is based on various static analysis methods such
as interval analysis, control dependence regions and information flow analysis. Furthermore, we
extend the concept of information flow analysis by allowing more expressive security policies. This
allows users to create more natural security policies similar to those that might be imposed on code
that inspects confidential data. We also give formal definitions of the theory used for the tool and
show how to extend these formal definitions. Moreover, we show the implementation of the tool
and how to extend the tool itself. The tool is capable of reasoning about complex eBPF programs,
with certain limitations. Such as missing support for some calls, not supporting byte or bit-specific
security levels, and using over-approximations for the interval analysis.

Supervisor(s)  

Thomas Jensen

External examiner(s)  

Carsten Elmar Schürmann

Date and time  

27-06-2023 at 13:30

Room  

A101 at HCØ

Name of student(s)  

Eva Kroon Enevoldsen

Study Programme  

Computer Science

Title  

DiCSG: Differentiable CSG Trees

Abstract  

Abstract
This thesis formulates a first try at a theory for defining differentiable constructive solid geometry
(CSG) trees. Automatically differentiable functions are part of the backbone of problem-solving in
both machine learning and general optimisation. CSG is a widely used tool, especially in computer
graphics and computer-aided design. The idea of CSG trees is to create complex shapes by combining
basic shapes, where the combinations are described in a tree structure. Developing differentiable CSG
trees allows CSG to enter the scene of optimisation and machine learning. This could enable CSG as
a tool for solving problems in even more contexts, as well as potentially bettering the performance of
CSG in its current use cases.
This thesis formulates how CSG trees can potentially be made differentiable using a general concept
of differentiability, enabling semi-smooth optimisation methods. First, a review of different shape
representations used in CSG is made, and then an analysis of how different parts of CSG trees can
be made differentiable is performed. Combining the review and analysis, the basics of differentiable
CSG trees are defined. From this, a theory is formulated to describe how to make the different parts
of CSG trees differentiable. This is done in two parts to cover two different operator sets commonly
used in CSG trees. For each theory set, a 2D prototype is implemented as a proof of concept for
differentiable CSG trees using the given operator set. Both prototypes successfully perform simple
optimisations, indicating that differentiable CSG trees can be created. The CSG trees of this thesis
are only differentiable in their shape and transformation parameters due to the non-continuous nature
of the other parameters defining CSG trees. An analysis is made on how to create a fully differentiable
CSG tree so it may be possible in future work.

Supervisor(s)  

Kenny Erleben

External examiner(s)  

Jakob Andreas Bærentzen

Date and time  

30-06-2023 at 13:30-15:00

Room  

Kenny's Office, DIKU building room 3.2.07.

Name of student(s)  

Matti Andreas Nielsen

Study Programme  

Computer Science

Title  

Add SLOG Single-Home & Multi-Home transactions to Distributed Snapper

Abstract  

According to the paper titled Serializable, low-latency, geo-replicated transactions(SLOG), geo distributed applications have been forced to give up at least one of Strict Serializability, low latency writes, and high transactional throughput. The paper presents ideas on how all these three things can be achieved by assigning regional homes to data items, implementing single-home and multi-home transactions using asynchronous state machine replication, and remastering data based on access heuristics. This thesis sets out to implement a prototype of SLOGs ideas on top of an already existing hybrid deterministic/non-deterministic transactions library known as Snapper. Snapper is built on
top of the platform Orleans, which is a cross-platform framework for building distributed applications using the virtual actor model. The value of the prototype have been assessed by comparing it to Snapper and the builtin transactions feature of Orleans. The thesis shows that this kind of
architecture can achieve up to 4x more throughput compared to Orleans transactions under varying contention levels, workload distributions and number of regions. The experiments also show that the prototype has significantly less throughput than desired when compared to Snapper, although the prototype has some additional benefits such as good locality on snapshot isolation reads. The experiments also show that this kind of prototype has a long way to go before reaching anything close to state of the art transaction throughput.

Supervisor(s)  

Yongluan Zhou

External examiner(s)  

Philippe Bonnet

Date and time  

07.07.2023 10:00

Room  

Sigurdsgade, 2-03

Name of student(s)  

Benjamin Paddags

Study Programme  

Computer Science

Title  

Automated Sentence Generation for a Spaced Repetition Software

Abstract  

This dissertation proposes and user-tests AllAI, an app that utilizes state-ofthe-art NLP technology to assist second language acquisition through spaced repetition, a procedure that spaces out exposure to each vocabulary item and thereby improves long-term recall. Other than current approaches, where words are either repeated solo and out of context, or fixed sentences are repeated, the proposed approach still schedules words independently but combines several words that are due for repetition into a dynamically chosen or generated sentence so that they are still learned in context. First, different NLP paradigms are investigated for their suitability to generate correct sentences that optimize the spaced repetition timing and it is found that retrieval using a BM25 ranking from a Wikipedia-based corpus, as well as a few-shot prompting approach both are suitable. Then a user study is carried out, comparing learning outcomes and user engagement in users using these two methods to the conventional approach of having a fixed sentence associated with each word. It was found that the use of the proposed sentence-based spaced repetition significantly increased learning outcomes (four- to six-fold) compared to
the conventional approach, primarily by increasing efficiency and vocabulary growth by showing more words more quickly, without decreasing the fraction of words remembered by learners. In the retrieval group, a significantly higher enjoyment was observed, possibly due to the higher efficiency, hinting at a higher user engagement.

Supervisor(s)  

Daniel Hershcovich and Valkyrie Savage

External examiner(s)  

Anders Jess Pedersen

Date and time  

22.08.2023 10:00

Room  

NLP meeting room in Vibenshus

Name of student(s)  

Sarah Howell

Study Programme  

Computer Science

Title  

Equitable Gaming: Substitution and Augmentation of Audio with Haptics for Deaf Gamers

Abstract  

Research question: which audio substitutions and augmentations can be integrated into video games to provide an equitable user experience for players who experience hearing loss? To explore this research question I designed, implemented, and tested a complete, arcade-style computer game titled Heartfelt Heroes. Heartfelt Heroes offered multiple
customizable haptic settings which were organized into four thematic groups. Hardof-hearing players were recruited to test out the game and play it four times, each playthrough having a different collection of active haptic settings.

After playing Heartfelt Heroes, participants filled out a user experience survey, then participated in an unstructured interview. The data gathered from surveys and interviews showed that audio augmentations via haptics tied to player actions were strongly preferred and increased player satisfaction, immersion, and sense of control. In addition, the data showed that users would prefer a standardized set of haptics to be implemented within similar genres, as well as a way to customize these haptic
settings in-game. The data also showed that players who experience hearing loss are unhappy with the current state of accessibility in video games and have a desire for the above improvements to be made.

This thesis establishes a foundation on which future audio augmentation and substitution implementations can be created. Consistently including haptic feedback on players’ immediate actions, creating standardized haptic settings, and in-game customization menus can be expanded in future research to create a more equitable user experience for individuals with varying degrees of hearing loss.

Supervisor(s)  

Pernille Bjørn and Valeria Borsotti

External examiner(s)  

Claus Witfelt

Date and time  

22.08.2023 13:00 - 14:00

Room  

Sigurdsgade 41

Name of student(s)  

Mikolaj Tymon Mazurczyk

Study Programme  

Computer Science

Title  

Estimating cosmic dust properties using Bayesian inference for data collected with the James Webb Space Telescope

Abstract  

Abstract
Cosmic dust plays a crucial role in the evolution of interstellar objects. Even
though its role in astronomical processes is relatively well understood, there
is still ongoing debate regarding its origin. The most likely sources of cosmic
dust in the universe are thought to be core-collapse supernovae. To confirm
that hypothesis, there is a need for accurate estimates of the quantities of
dust ejected by a dying star. The dust formed around supernovae is measured
through the electromagnetic radiation that they emit, which can be captured
by advanced telescopes, such as the James Webb Space Telescope (JWST).
Recently, there have been attempts to apply machine learning to test this hypothesis. Since real-world data is scarce, models are trained with a simulated
dataset of JWST observations of supernova explosions. However, optimizing regular models predicting the dust features based on JWST observations
displayed several limitations.
For instance, collecting data with JWST is resource-consuming, and since
not all of its instruments can be used at once, this implies that, typically, there
is missing data. Furthermore, many of the dust features turn out to be nearly
impossible to predict accurately due to their complex distributions. Hence,
there is a need for robust estimates of uncertainties around the predictions.
The approach presented in this thesis aims to address these challenges
through the application of Bayesian inference to the issue. I propose to represent the Bayesian likelihood density with the deep learning model. Subsequently, employing the Markov Chain Monte Carlo technique allows me to
sample from the posterior. Choosing the distribution predicted by the model
so that one can easily marginalize filters from it addresses the first limitation.
Furthermore, the representation of the distribution of dust features with the
Bayes rule allows for an intricate and accurate estimate of the posterior. Additionally, this work offers a comparative study between a regular neural network
and a more advanced architecture, the Conditional Variational Autoencoder
(CVAE), in their ability to represent likelihood density.
In conclusion, my results prove that CVAE does not provide significantly
better results than the regular neural network. Nonetheless, both still performed better than the non-Bayesian standard method, even with some of
the JWST observations marginalized, achieving the primary objective of the
thesis.

Supervisor(s)  

Oswin Krause

External examiner(s)  

Melih Kandemir

Date and time  

25 September 2023 at 9:00-10:00

Room  

UP1-2-0-04

Name of student(s)  

Jonas Masiulionis and Xinzhi Huo

Study Programme  

Computer Science

Title  

The Satisfaction of Movement Systems in Video Games

Abstract  

In this study, we have investigated how specific movement design decisions affect a user’s video game enjoyment. The primary aim was to find out if there is a clear preference towards aspects of the movement system. Our test specifically focused on the 2D platformer genre. We created
multiple modules based on popular platformer mechanics and had participants test each possible combination, giving their thoughts on them after playing each combination. We have found that gliding mechanics, with time trial challenges and fast theming should be preferred in a 2D
platformer. While, dashing mechanics, tracing challenges and slow theming negatively impact the player’s opinion about the game.

Supervisor(s)  

Valkyrie Arline Savage

External examiner(s)  

Louise Petersen Matjeka

Date and time  

29.09.2023 09:00 - 10:30

Room  

Sigurdsgade 0-11

Name of student(s)  

Sandra Martinková

Study Programme  

Statistics

Title  

Analysis of gender bias methods in NLP for West Slavic languages

Abstract  

Pre-trained language models have been known to perpetuate biases from the underlying datasets to downstream tasks. However, these findings are predominantly based on monolingual language models for English, whereas there are few investigative studies of biases encoded in language models for languages beyond English. In this work, we fill this gap by analysing gender bias in West Slavic language models. We introduce a template-based dataset in Czech, Polish, and Slovak for measuring gender bias towards male, female and non-binary subjects. We complete the sentences using both mono- and multilingual language models and assess their suitability for the masked language modelling objective. Next, we measure gender bias encoded in West Slavic language models using lexicon-based gender bias evaluation, in particular HurtLex lexica, HONEST score and Valence, Arousal and Dominance lexica. We find that these language models produce hurtful completions that depend on the subject’s gender. Perhaps surprisingly, Czech and Slovak language models produce more hurtful completions with men as subjects, which, upon inspection, we find is due to completions being related to violence, death, and sickness. For the Polish model, most hurtful completions are generated
with female templates and there’s a large number of completions related to prostitution. Additionally, we find that male templates have more extreme completions in 75% of our experimental settings, i.e. completions such that these words have either a very large or a very small valence, arousal or dominance scores.

Supervisor(s)  

Isabelle Augenstein

External examiner(s)  

Niels Væver Hartvig

Date and time  

17.07.2023 15:00 - 16:00

Room  

Room 01-1-223, Øster Voldgade 3

Name of student(s)  

Zhuyang Lin

Study Programme  

Bioinformatics

Title  

Multimodal Deep Generative Decoder

Abstract  

Abstract
With the advancement of machine learning, the capacity of models to ingest and
comprehend multimodal data has grown increasingly significant. This development
not only applies to large models such as GPT-4 and PaLM-E but also plays a pivotal
role in other small models like VAE. In this research, we have implemented a
Multimodal Deep Generative Decoder (MDGD), which is able to learn from clinical
data that is characteristically heterogeneous and sparse. To manage the issue of
missing data, we also implemented a Mask to handle the missing data and perform
missing data imputation. Furthermore, to alleviate the posterior collapse problem,
we improved the MDGD by integrate weights on the model. Our experimental results
demonstrate that the modified MDGD exhibits commendable performance in data
reconstruction as well as missing data imputation within the clinical dataset. The
findings of this study illuminate the potential of DGD in multimodal data learning
and also propose an efficacious method for augmenting the original DGD model.

Supervisor(s)  

Anders Krogh

External examiner(s)  

Ole Lund

Date and time  

13-06-2023 at 10:30

Room  

Seminar room, Panum 33.4.D

Name of student(s)  

Michaela Válková

Study Programme  

Bioinformatics

Title  

Supervised Machine Learning in Migraine

Abstract  

Abstract
Migraine is a common neurovascular disorder affecting over a billion people worldwide,
predominantly women. However, the vast heterogeneity of phenotype characteristics among
patients makes diagnosis challenging. While vasoconstrictor triptans are the most commonly
used acute treatment, they only provide pain relief in 60-70% of patients, and the mechanisms
underlying this variability in response among individuals are not yet fully understood.
The aim of this project is to detect the relationship between phenotypic features and migraine
acute treatment response using supervised machine learning. Data is derived from a Danish
migraine population cohort, a subgroup of the Danish Blood Donor Study, comprising approximately 2500 triptan users with detailed phenotypic information gathered from a diagnostic
migraine questionnaire. The primary method used is Qlattice, a novel approach to symbolic
regression, that emphasizes model interpretability. Additionally, random forest is utilized as
an established method to provide baseline comparisons.
The findings confirm previously observed correlations between migraine aura and treatment
failure, and suggest an additive effect of visual and sensory aura. Furthermore, presence of
nausea and bilateral headache shows to further reduce the probability of successful treatment.
Finally, the results indicate that the probability of successful treatment increases with age.
The QLattice model produced consistent results on the testing dataset with an area under the
receiving operator curve of 0.65, which was higher than that of random forest. The importance
of the features detected by QLattice was confirmed with random forest.
The findings presented in this study can potentially contribute to understanding the triptan mechanisms and to development of precision medicine for migraine, providing a more
tailored approach to treatment based on an individual’s specific phenotypic characteristics.
Furthermore, this project can serve as a demonstration of the use of QLattice with clinical
questionnaire data.

Supervisor(s)  

Christian Igel

External examiner(s)  

Jose MG (Txema) Izarzugaza

Date and time  

22-06-2023 at 10:30

Room  

Online

Name of student(s)  

Zuzanna Maria Dubanowska

Study Programme  

IT and Cognition

Title  

Summarisation with pixels

Abstract  

Abstract
The internet has revolutionised how we create, publish and access content offering an
instant supply to limitless knowledge online. Finding relevant information, though,
proves challenging amidst this sea of data, thus it became crucial to find ways to
easily identify, extract and condense valuable knowledge. Automatic text summarisation (ATS) is the process of generating short, informative accounts of larger texts
using techniques from Natural Language Processing, Machine Learning and statistics without human help. It has potential to improve every area where we encounter
unstructured textual data, from education [1], academia [2], the legal industry [3, 4]
to healthcare [5]. We distinct two approaches to ATS: extractive and abstractive.
In the former, the summary is a concatenated collection of vital sentences from the
original document. In the latter, the system generates new text which paraphrases
the sense of the original document.
Since the advent of the Transformer [6], Large Language Models have been continually improving state-of-the-art in abstractive text summarisation, however research
efforts are still needed to bridge the gap between human- and machine-generated
summaries. Vocabulary bottlenecks are a known issue LLMs suffer from, which
can hinder their performance in summarisation tasks as finite vocabularies limit the
amount of words a model can use to generate text.
In this work, we investigate the potential of Pixel-based Encoder of Language
(PIXEL) [7] for text summarisation. PIXEL frames language modelling as a visual recognition task and operates on visual text representations instead of discrete
tokens, mitigating the bottleneck issues. We propose PIXELSum, an abstractive
summarisation Seq2Seq model with a pre-trained PIXEL encoder and pre-trained
GPT2 decoder. We examine how using different sizes / architectures of the decoder
and freezing different parts of the network affects performance when fine-tuned on
a summarisation dataset.
The PIXELSum variant where GPT2-medium decoder was used performs the best,
with two best performing models depending on the metric: GPT2-medium where
the whole decoder was fine-tuned, achieving a ROUGE1 of 16.23, and GPT2-medium
where only cross-attention was fine-tuned, achieving a BERTScore precision of
61.19%. The generated text for all model variants is fluent, but factually inconsistent and irrelevant, injecting information not present in the document.
Our experimental results give rise to a number of observations, shedding light on
what is the most significant in PIXEL-based summarisation models. Evidence shows
that when updating the weights of cross-attention only, the model learns steadily, in
other cases it overfits. There are indications that larger decoders (300M+ params.)
were over-parameterised for the task. Evidence points to that the decoder did not
know how to attend to encoder outputs. Taking these observations together, we
suggest, that to leverage PIXEL for text generation successfully, the full resulting
Seq2Seq model should be pre-trained on a generalised summarisation-alike task,
then the cross-attention layers fine-tuned on specific summarisation tasks.
Keywords: Language Modelling, Text Summarisation, Visual Text Representations

Supervisor(s)  

Desmond Elliot

External examiner(s)  

Claus Witfelt

Date and time  

12-06-2023 at 10:00-11:00

Room  

Vibenhuset 4th Floor Meeting Room

Name of student(s)  

Zhaoqi Liu

Study Programme  

IT and Cognition

Title  

Systematic Study of Robustness within Natural Language Inference Models

Abstract  

A B S T R A C T
Recent years have witnessed an explosive increase in transformer-based large language models
(LLMs). Many of them have demonstrated remarkable performance on a variety of downstream
tasks and even zero-shot generalization in the field of natural language processing (NLP), which
is widely attributed to the comprehension of world knowledge. However, the semantic robustness
of these models remains questionable. Focusing on the natural language inference (NLI) task, this
study proposes a novel evaluation pipeline for the robustness test of NLI systems and provides a
systematic analysis scheme targeted at models’ behavior in in-domain performance and out-of-domain
generalization. Through the experiments, I find that the NLI models are not semantically robust and
fail to form a self-consistent inference logic.

Supervisor(s)  

Isabelle Augenstein

External examiner(s)  

Zeljko Agic

Date and time  

19-06-2023 at 13:00-14:00

Room  

Room 01-1-223 at Østervold Observatory

Name of student(s)  

Xinyue Hu

Study Programme  

IT and Cognition

Title  

Investigating the Embodied Experience in VR using a Micro-Phenomenological Method: What Happens in the Moment of Realignment?

Abstract  

Abstract
In this project, our focus lies in exploring the sense of embodiment that arises from virtual
reality (VR) experiences. Through the utilization of micro-phenomenological methods, we aim
to gather qualitative e data to gain a comprehensive understanding of the cognitive aspect
related to owning a virtual body. Our ultimate goal is to understand the cognitive experience of
being in the virtual world and what happened in the moment of realignment from the real
world to the virtual world.
As a preliminary step, we conducted a preparation project to assess the feasibility of
extracting the shared structure of the virtual body ownership experience with a qualitative
method. This was accomplished by employing simplistic avatars and environments to develop a
semi-structured interview guide.
The concept of "embodiment," as described by Kilteni et al. (2012), encompasses three
key elements: self-location, agency, and body ownership. Numerous studies have explored the
phenomenon of the body-ownership illusion using morphologically altered virtual avatars.
However, the overall encoding of participants' experiences and the representation of these
virtual interactions in their minds remain unclear. The micro-phenomenological approach
allows us to gather extensive data on participants' internal narratives of their experiences,
enabling us to gain insights into the induction of the illusion. Ultimately, our objective is to
identify a shared representation and structure of the body-ownership illusion among
participants by analyzing how they encode information in a virtual environment and
communicate the embodied experience and compare these results to the theoretical framework
of embodied experience in the virtual environment.

Supervisor(s)  

Kasper Hornbæk

External examiner(s)  

Jakob Eg Larsen

Date and time  

28-06-2023

Room  

Sigurdsgade

Name of student(s)  

Yuqin Zhou

Study Programme  

IT and Cognition

Title  

Tensor Networks for Language Modeling

Abstract  

Abstract
Modern neural networks have shown success in language modeling tasks. Tensor networks
(TNs) have been used to unravel their theoretical properties, albeit without concrete empirical support: no TN-like language model has proven effective on widely accepted natural
language processing datasets. This thesis fills this gap by presenting a class of tensor-train
language models (TTLMs). TTLMs encode the joint probability distribution of sequences
into a wave function and operate on conditional probability distributions during training and
inference. Theoretically, we demonstrate that the architectures of Second-order Recurrent
Neural Networks (RNNs), Recurrent Arithmetic Circuits, and Multiplicative Integration RNNs
are essentially special cases of TTLMs. Experimental evaluations reveal that the theoretical
properties of TNs—linearity, multiplicativity, and complex-valued representations—cause difficulties during training. We employ statistical methods, such as the Kolmogorov-Smirnov test,
to identify that the issues are unstable gradient flows, exponential decay in hidden states, and
complex number multiplications. To enhance the effectiveness of TTLMs, we propose a class
of variants called the Linear Multiplicative Models (LMMs). LMMs are linear, multiplicative,
and complex-valued, achieving competitive results on language modeling tasks compared to
vanilla RNNs. Their success provides empirical support for prior research that establishes
the equivalence between TNs and neural networks, and offers a novel perspective in the field
dominated by nonlinear and additive language models.

Supervisor(s)  

Jakob Grue Simonsen

External examiner(s)  

Christian Hardmeier

Date and time  

6 October 2023 at 9:15

Room  

DIKU - Room 2-0-0-4