MSc Defences Spring 2023

See the list of MSc thesis defences at DIKU this spring. The list will continuously be updated.

Information about the thesis, supervisor, location of the defence, etc. can be found on the respective events below.

Computer Science

 

 

 

 

 

 

Name of student(s)  

Nicklas Christiansen

Study Programme  

Computer Science

Title  

Automatic recommendation of players in the Danish Superliga using Twitter posts and simple player data

Abstract  

Football clubs are spending millions each year recruiting new players. This thesis investigates the potential in using Machine Learning to help the recruitment process in the different clubs. The data used is simple data about the players and teams, who have played in the Danish Superliga since 2019. Twitter is scraped for posts about the players to enhance the feature set. Entity-level sentiment analysis is performed on these posts using pretrained models and the process is carefully validated. A general problem throughout this thesis is the lack of interactions with the players from various different users, but scouts and professionals in Brøndby IF have labelled a small data set (800 interactions) for experimentation
by marking players interesting, not interesting or neither. Standard evaluation metrics (MRR, MAP, TopK-accuracy and TopK-hit rate) are used and the most complicated and promising model hits an MRR of 1 and MAP of 0.85, when evaluating the top3 retrieved players with the forementioned scouts.

Supervisor(s)  

Christina Lioma

External examiner(s)  

Jes Frellsen

Date and time  

20.03.2023 13:00 - 14:00

Room  

Room 1.2.26 at DIKU

 

 

Name of student(s)  

Philip Meier Kreutzer

Study Programme  

Computer Science

Title  

Introducing programming constructs for ease of implementing cryptographic algorithms in Hermes

Abstract  

In this thesis, we will be investigating the Hermes reverisble programming
language. The language and programming model shows promise and provides guarantees that are useful for implementing encryption algorithms. We will be investigating ways of improving the language, to allow for easier and better implementation of encryption algorithms.

Currently the language has some barriers making implementation of algorithms using S-boxes exceedingly difficult. We will be looking for ways to remediate this. To do this, we have looked into a way to do logical synhtesis for reversible programming. Furthermore, we have showed a way to extend the grammar of Hermes to introduce constructs, to help implement S-boxes. This is done in a way, where the grammar looks much like a constant lookup table has been defined.

The Changes to the language, through the implementation of a logical synthesis algorithm, has made encryption algorithms utilizing static S-boxes, simpler to implement in Hermes.

Supervisor(s)  

Michael Kirkedal Thomsen

External examiner(s)  

Mads Rosendahl

Date and time  

24.03.2023 11:00

Room  

PLTC meeting room

 

Name of student(s)  

Yuan Chen

Study Programme  

Computer Science

Title  

Event Manipulation Language, Modeling a Domain-Specific Language for Managing Explicit Causal Inconsistencies in Asynchronous Event-based Distributed Systems: A Microservice Architecture Approach

Abstract  

The main objective of this thesis project is to design and implement an event manipulation language for databases in microservice architectures that can act upon detected data consistency violations, as defined by the application developer. This language should be lightweight, flexible, and designed to work with the dynamic monitoring system MonPoly. Further, a secondary goal is to construct an exemplary microservice system on top of the Orleans framework [7] for bench-marking the functionalities of the language.

Supervisor(s)  

Boris Düdder (primary supervisor) and Tilman Zuckmantel

External examiner(s)  

Pavel Hruby

Date and time  

24.03.2023 10:00

Room  

Room 2.03 in Sigurdsgade 41

 

Name of student(s)  

Troels Korreman Nielsen

Study Programme  

Computer Science

Title  

Verification Condition Generation for eBPF

Abstract  

The eBPF subsystem for Linux allows loading programs at runtime from user space into kernel space for execution. Using static analysis and JIT-compilation, both high performance and safety can be guaranteed. However, there is a conflict of interest between the desire to expand the
capabilities of the eBPF program verifier and the need to manage its complexity. Proof-carrying code (PCC) is an alternative approach to program validation. It shifts the burden of validating
safety to the code producer by requiring that a safety proof be submitted with every program. By replacing the static analyzer with a proof checker, consumers reduce their workload while producers can submit a wider range of programs. In this project, the prospect of using PCC to validate eBPF is explored with a focus on branching and looping programs. Subsets of the eBPF instruction set and safety policies are chosen as targets. Verification condition generators (VC-gen) for partial and total correctness are designed for these targets. A proof of concept VC-gen for partial correctness is implemented and compared to the eBPF verifier on synthetic programs. By dispatching verification conditions to a solver, PCC is shown to be able to validate several safe programs that the verifier rejects.

Supervisor(s)  

Ken Friis Larsen

External examiner(s)  

Carsten Elmar Schürmann

Date and time  

27.03.2023 15:15

Room  

PLTC meeting room at HCØ

 

Name of student(s)  

Jonas Østergaard Klausen

Study Programme  

Computer Science

Title  

Peelability and Tabulation Hashing

Abstract  

Simple tabulation is a hashing scheme based on memory lookups and fast xor operations, originally presented by Zobrist [ICGA Journal 1990] for efficiently storing the state of games. The scheme is only 3-independent but will produce independent hash values as long as the input key set is devoid of a certain kind of structure.

A sufficient condition for a set of keys to be hashed independently is for them to be peelable. A family of sets is said to be peelable if sets can be removed one at a time, each containing a unique character not found in the remainder of the sets.

In this work we present a new proof that simple tabulation h : ϕc → Σd
will produce a peelable set of hash values with high probability when applied to a set of at most |Σ|/2 keys. A more general statement was shown in the work of Dahlgaard, Knudsen, Rotenberg and Thorup [FOCS 2015] but our proof is simpler and free from both asymptotics and hidden constants. Our work is thus the first to give concrete guarantees on the validity of these results for fixed parameters.

This immediately gives results on the independence of double tabulation [Thorup, FOCS 2013], where two simple tabulation functions are composed, thus applying simple tabulation to the set of now-peelable keys. Peelability is also the essential ingredient for invertible bloom filters [Eppstein, Goodrich, TKDE 2011], a simple randomized data structure for storing a set of elements through the use of a hash function. Our result thus shows that simple tabulation is a suitable scheme for this data structure.

Supervisor(s)  

Mikkel Thorup

External examiner(s)  

Eva Rotenberg

Date and time  

28.03.2023 10:00

Room  

Online

 

Name of student(s)  

Mikkel Brok Reiter Sørensen

Study Programme  

Computer Science

Title  

Creating a Pneumatic Silicone-based Stewart Platform

Abstract  

The aim of this thesis is to explore options for creating a 3D printable Stewart platform, delimited by six degrees of freedom (DoF) for use in magnetic resonance (MR) scans. This imposes certain restrictions on material usage; in particular, there must be no ferromagnetic parts within the scanner, and in general the amount of metal components must be limited in order not to interfere with imaging. As part of the process, printed circuit boards (PCB’s) are designed and manufactured, while a microcontroller is programmed to control the platform via an API on a host computer. As the platform should be able to operate inside an MR scanner, the only feedback on platform position and orientation is from air pressure sensors, located a sufficient distance from the scanner.

The thesis explores different methods of manufacturing pneumatic actuators cast from silicone. The process includes attempts at creating embedded constraining layers using 3D printed elements, to better control the actuators.

After manufacturing a proof of concept of the platform, tests are performed to validate whether the platform is within specifications, which are defined by available data describing head movement on patients in MR scanners.

Overall, this thesis describes the potential of silicone-based pneumatic actuators as part of a 3D printed platform, as well as challenges with sensing movements as a result of having a pneumatic core of a load-bearing platform.

Supervisor(s)  

Kenny Erleben

External examiner(s)  

Jeppe Revall Frisvad

Date and time  

31.03.2023 10:00

Room  

DIKU UP1, lokale 3.2.07

 

 

Name of student(s)  

Aleksa Durdevic

Study Programme  

Computer Science

Title  

Distributed Lock Manager, Providing serializability to distributed transactions

Abstract  

Microservices have become a popular alternative to monolithic systems, providing benefits such as scalability, loose coupling, and fault tolerance. However, ensuring the ACID properties required for critical transactions in distributed systems can be challenging. This is especially true when transactions require coordination between multiple microservices. To address this, we propose the development of a distributed lock manager that utilizes multiple granularity locking. The lock manager will be implemented as a microservice and will provide serializability for transactions by locking shared resources appropriately. Our lock manager will work using a hierarchical tree structure, where each node represents a specific lockable resource, and will use a unique identifier for each
transaction and lockable item. The implementation will consist of three microservices: the lock manager, a client microservice, and a store microservice. The store microservice will contain resources shared by multiple transactions, while the client microservice will utilize the lock manager to execute parallel transactions. Our proposed lock manager will operate in an environment with a high number of concurrent transactions often requesting access to the same resources.

Supervisor(s)  

Boris Düdder and Tilman Zuckmantel

External examiner(s)  

Pavel Hruby

Date and time  

14.04.2023 10:45

Room  

Room 2.03 in Sigurdsgade 41

 

 

Name of student(s)  

Niklas Magnussen

Study Programme  

Computer Science

Title  

Multiplanar Serial Classification

Abstract  

Knee osteoarthritis is degenerative disease that affect a large percentage of older adults. The disease leads to pain, decreased mobility and in most cases the effect on the patients’ quality of life leads to the knee having to be replaced. Early detection of whether patients will require a total knee replacement can help medical professionals make more informed decisions about the patient’s care and treatment options. By identifying patients who are at higher risk for total knee replacements, doctors can adjust the choice of treatments, such as physical therapy or drugs, to delay or avoid the need for surgery. In this thesis we propose an approach to total knee replacement prediction using serial MRI data, which is based on the multiplanar strategy of the multiplanar U-net for segmentation tasks. We wished to investigate whether the computational benefits and statistical robustness of the multiplanar strategy will carry over when moving from a segmentation task to a classification task, and we compare the multiplanar classification method to a different imaging-only deep learning approach, which uses imaging biomarkers computed from segmentation volumes of the knees. The results of the trained models show the multiplanar classification methods perform similarly or better than the imaging biomarker methods, with much opportunity for further researching and developing the multiplanar strategy for the task of classification.

Supervisor(s)  

Erik Bjørnager Dam

External examiner(s)  

Dan Witzner Hansen

Date and time  

28.04.2023 14:00

Room  

UP1-2-0-04 / UP1-2-0-06