MSc Defences Summer 2022

Study Programme

Computer Science

Title

Generating Generators - A CLS-based Pipeline for Unity

Abstract
This project aims to provide a prototyping pipeline for game development with the focus of using procedural generation. Inspiring game and level designers to develop game ideas with a foundation in procedural generation. Furthermore, the prototyping tool provides a high-level user interface allowing the game and level designers to generate prototypes
without having to consult programmers that must spend time on prototypes that may or may not be used. The pipeline uses the CLS framework by Bessai et. al. ([1]) for code generation and features Unity integration.

Supervisor

Boris Düdder

External examiner

Mads Rosendahl

Time and place

10:00 at Sigurdsgade 41, room 2.03 

Study Programme

Computer Science

Title

Distance Oracles: Exploring the Pattern Approach for Exact Distances in Planar Graphs

Abstract

We give a survey on the literature on distance oracles spanning exact and approximate distance reporting in the setting of both general and planar graphs.
Given an undirected, unweighted planar graph, Mozes, Fredslund-Hansen and Wulff-Nilsen gave the first truly sub-quadtric size oracle answering exact distances in constant time. We give a thorough description of this oracle, contributing with an analysis of quadratic preprocessing time and a new analysis of space requirement, achieving their stated result while cor-recting the inadequacies of the original paper.

A key element in achieving this result lies in bounding the number of distinct discrete derivatives w.r.t. the distance between a source vertex and vertices on a cycle, commonly referred to as patterns. It has been proven that a cubic upper bound exists, however it is hypothesized that a quadratic upper bound is likely true. We give a quadratic lower bound and strengthen the quadratic hypothesis through experimentation. In addition, we give new analysis predicting the performance of the oracle above assuming the quadratic hypothesis true, yielding stronger guarantees than previously known.

A second key element is to decompose the input graph into r -divisions, for which we give a much simplified scheme based on tree regions. Given an n-vertex planar non-triangulated graph G and a parameter 0 < r ≤ n, our scheme decomposes G into Θ(n/r ) connected regions of size Θ(r ) vertices, each with just 1 hole, in linear time. This matches the complex decom-position scheme by Mozes, Klein and Sommer except for a regionboundary guarantee, not required for some applications including this oracle.

Finally we show an implementation of the oracle using our simplified decomposition scheme. Through experimentation we thus provide evidence of a truly sub-quadratic distance oracle answering queries in constant time, straight forward to implement and without showing expensive constants invalidating applicability.

Supervisor(s)

Christian Wulff-Nilsen

External examiner(s)

Jesper Larsen

Time and place

13.06.2022 09:30

Lille UP1 på DIKU, Universitetsparken 1

Study Programme

Computer Science

Title

Multi-task learning on images of beetles using deep neural networks

Abstract

Deep learning image classification is usually performed at only one classification level. However, explicit hierarchical information about the objects being classified is rarely leveraged. We investigate ways of leveraging such hierarchical information, concretely in the context of classification of images of rove beetles. We propose a novel multi-task classifier, the Fusion model. The fusion model is characterized by 1) a number of tracks corresponding to the hierarchical levels used for classification and 2) a method by which information can be shared across said tracks. Thereby the fusion model allows flexible parameter sharing between different classification tasks. We also propose an extension to the fusion model, which incorporates a reconstructive task. We hypothesise that including a reconstructive task will be amenable to classification task performance and vice versa. However, we cannot report conclusive evidence for or against the utility of these proposed models. Experiments indicate that there may be fundamental issues with the dataset of rove beetles. In particular, we observe in our analysis that independently trained models consistently misclassify specific samples from the rove beetle dataset, which we confirm to be mislabelled samples. In future work, we wish to investigate to what extent samples from the rove beetle dataset may be mislabelled. All our code is publicly available at https://github.com/pants007/msc-thesis-code. 

Supervisor(s)

Kim Steenstrup Pedersen

External examiner(s)

Rasmus Reinhold Paulsen

Time and place

13.06.2022 10:30 - 12:15

Zoo Museum, Universitetsparken 15, lille mødelokale

Study Programme

Computer Science

Title

Meaning Representation of Legal Documents

Abstract

AMR has shown to be useful for many downstream tasks such as Common Sense Reasoning and Information Extraction. In this work, I explore the use of AMR for legal and logical reasoning. Specifically, I am proposing four model architectures that utilize semantic encoding in the form of AMR. The AMR graphs are generated by using an AMR parser. The presented architectures use graph linearization in combination with pre-trained language models (PLM). The PLMs are Legal-BERT, which was pre-trained on legal text, and AMRBART, which was pre-trained on AMR graphs.

Additionally, I propose an architecture that combines AMR and textual input by concatenating text and AMR embeddings extracted from PLMs. The results show that the AMR models are not able to outperform text-only baseline models. Although the performance could not be improved, it was shown that the AMR models can solve different instances than the text models. The error analysis further revealed that AMR parsing quality is the biggest problem, especially regarding inputs with multiple sentences. Furthermore, a theoretical analysis of how logical relations are represented in AMR has shown that AMR might be helpful to capture logical relations in some logical statements while it might not be useful for others. 

Supervisor(s)

Daniel Hershcovich

External examiner(s)

Christian Hardmeier

Time and place

14.06.2022 14:00

Online

Study Programme

Computer Science

Title

Multilingual Information Retrieval with word embeddings

Abstract

Multilingual Information Retrieval (MLIR) is a task of searching and ranking relevant documents from a collection written in various languages, given the user’s query. With easy access to multilingual resources, there are many use cases of the problem, e.g., searching for patents, medical treatments, foreign news, or culture. On top of the challenges coming from searching the monolingual collection, MLIR faces the challenge of translation and creating a ranking that is not groundlessly biased
towards any language. A solution to these problems is offered by multilingual word embeddings which proved to achieve high performance in many language understanding tasks. In this thesis, I explore multilingual BERT word embeddings in the document ranking task for the multilingual IR. Additionally, I investigate the impact of the embedding space alignment to remove the language-specific characteristics and allow better document comparison and, as a result, better final ranking.

The findings show that zero-shot training works well in the MLIR setting, even if the pre-trained models are developed using only English data, and significantly improve the performance compared to the baseline. Additionally, the vector space alignment improves the language distribution over the top-ranked documents. 

Supervisor(s)

Maria Maistro

External examiner(s)

Luca Maria Aiello

Time and place

14.06.2022 10:30

3-0-25 at DIKU, Universitetsparken 1

Study Programme

Computer Science

Title

Visual growth patterns in Metarhizium

Abstract
Deep learning models are used for an increasing variety of purposes. With capabilities such as semantic segmentation, they are ideal for tasks in which objects need to be localized and measured. Deep learning models are however restricted by the requirement of large datasets for training.

The Department of Plant and Environmental Sciences at the University of Copenhagen (PLEN) has acquired the Reshape Imaging System (RIS) which is a machine capable of imaging petri dishes at set intervals. This creates large amounts of image data which can then be used to develop such deep learning models for analysis. To train models, true labels are needed. These can be created in an intelligent manner employing active learning.

In this thesis, the possibility of applying deep learning models to growth analysis of the fungi Metarhizium is investigated. Two experiments were conducted on the RIS, where Metarhizium was inoculated on petri dishes in the center and multiple places respectively. A pipeline for model development regarding such experiments is suggested. U-net models were developed for semantic segmentation and produced good results on both experiments. The first U-net produced segmentations on a test set with an average Dice score of 0.9647. For the second experiment Monte Carlo Dropout was applied as a way to increase the efficiency of the manual annotation procedure. This experiment shows that models can be trained with fewer annotations when the unlabelled data is selected intelligently for manual annotation. This segmentation task was harder in nature, as the fungi were located in differing areas of the petri dishes. The final U-net segmentation model achieved an average Dice score of 0.8471 on a test set. This score can largely be excused by the nature of the Dice score. In the start of the growth phase, when the fungi first appear, the model is usually off by a few frames, producing segmentations with a Dice score of 0 in these instances dragging the average Dice score down.

In conclusion deep learning models are capable of measuring the growth of Metarhizium via image data from the RIS.
The production of manual annotations for training data can be made more efficient with Monte Carlo Dropout.

Supervisor

Jon Sporring

External examiner

Rasmus Reinhold Paulsen

Time and place

Small UP1 at DIKU

Study Programme

Computer Science

Title

Improving the User Experiences in VR by Affecting Time Perception With High Social Density

Abstract
The term user experience describes user perceptions and responses that result from the use of a product. Losing track of time is a commonly reported experience and an indication of a good user experience (UX).
Therefore, there might be potential relationships between user experience and time perception.

The ideas behind this work arise from threefold elements. 1. This work assumes that better user experiences relate to the phenomenon of time underestimation or faster self-perceived time speed; 2. A previous study showed that participants underestimated the duration of video clips with high social density and overestimated the duration of video clips with low social density in virtual reality (VR) [114]; 3. The effects of social facilitation/inhibition caused by the presence of others will influence individual behavior and performance [13], and the effects will be stronger if there are more people [56, 64]. Thus, this work attempts to use social density as an independent variable and asks participants to estimate seven minutes while playing a bubble-hitting task in VR. In this work, different social density means a different number of virtual agents in the scenarios.
This work presents the hypotheses in three aspects: 1. High social density will lead to time underestimation or faster self-sensed time speed; 2. High social density will lead to better user experiences, which could be revealed by higher scores on UX-related questions; 3. High social density will lead to stronger social facilitation/inhibition effects, which will make the easy task (bubble hitting task) easier and make the hard task (timing task) harder.
Results of this work against the hypothesis of shorter time estimation in high social density conditions, but support most of the UX-related hypotheses and almost all the hypotheses associated with social
facilitation/inhibition effects. Specifically, the results show that participants spent more time in the high social density condition (mean 437.38 seconds) than in the low social density condition (mean 403.25 seconds) until they thought they had played for seven minutes. However, there was a trend that participants had a faster self-sensed time speed under high social density conditions. Furthermore, participants had significantly higher scores on involvement and overall satisfaction, while they had a significantly lower score on the sense of mastery under conditions of high social density. According to environmental psychology studies, for a timing task as in this work, people are expected to spend more time in a comfortable environment, that is, they will spend more time in a comfortable environment before they feel it is time to stop [38]. Longer time estimation and faster self-sensed time of high social density conditions in this work imply that people may think quiet VR environments with high social density are more comfortable than the same environments with low social density. The results also provide guidelines for future VR application designers, stating that they can improve users’ involvement
3 and overall satisfaction by increasing the number of virtual agents in the virtual environment. In addition, the results indicate a new way to adjust the self-sensed difficulty in applications such as games without
adjusting the objective difficulty, that is, increasing the social density of the scenarios. Limitations and future work are discussed in the last part of this report. In all, high social density in a quiet VR environment leads to significantly higher involvement and overall satisfaction, while significantly reducing the sense of mastery. These may be caused by social
facilitation/inhibition effects. Future VR applications should consider increasing the number of virtual agents in environments to improve user experiences and self-sensed difficulty.

Supervisor

Kasper Hornbæk

External examiner

Henrique Galvan Debarba

Time and place

9:30 

Place: TBD

Study Programme

Computer Science

Title

Probing Pre-trained Langugage Models for Cross-Cultural Differences in Values

Abstract

The ubiquity of Language Models (LMs) in Natural Language Processing research and the daily lives of people around the world calls for a more thorough understanding of their embedded biases and behaviour in different contexts. Further, language we use often reflects our social, cultural and political values and since LMs are trained on web text, they pick up on these values. In order to investigate and better understand this phenomena, I present the first study, to the best of my knowledge, to surface bias towards values across different cultures picked up by pre-trained LMs. I develop a framework for probing three widely used multilingual LMs namely mBERT, XLM and XLM-R by creating probes in 13 languages based on questions asked in two wide ranging value surveys. The surveys were developed by social scientists and psychologists in order to analyse values across different cultures and there is a large body of work in disciplines within the social sciences using results from the surveys to get insights and develop theories. Through my experiments, I find that each of three models capture differences in values across cultures. There are, however, discrepancies in the values picked up by the LMs across different cultures. They vary in range as well as have low alignment with each other in terms predicted value scores. I also find that the predicted scores for values across cultures do not correlate with survey data filled in by people from those cultures. I present a discussion around the implications of this misalignment when using models in a cross-cultural context and suggest ways for bridging the gap.

Keywords: probing, language models, values, cross-cultural

Supervisor

Isabelle Augenstein

External examiner

Zeljko Agic

Time and place

12:50 -14:10

UP1-1-1-N116B, DIKU, Universitetsparken 1, 2100 CPH Ø

Study Programme

Computer Science

Title

Evaluating Parsers with Chinese Challenge Set

Abstract

In dependency parsing and abstract meaning representation (AMR) parsing, there may be several grammatically correct ways to parse a sentence. Considering the meaning of the sentence can help disambiguation, as there may be only one semantically correctly way to parse. The ability to disambiguate is important for the parsers since many downstream tasks including machine translation require clear and unambiguous results. To assess parsers in such situations, we proposed a Chinese challenge set focusing on four sentence structures that involve ambiguities resolved by selectional preferences and commonsense reasoning. We investigated some well-performing dependency parsers from CoNLL 2018 shared task (Zeman et al., 2018) in depth and evaluated some of them on our challenge set, showing that each parser is biased differently, and that stanza is doing the best while UDPipe1 is the worst. We also find out that tokenization is a major cause of deterioration on parsing performance for Chinese parsers. We further evaluated these parsers with general Chinese, English and German test set, the result shows the relative performance between parsers is consistent throughout different languages, and Chinese is the most confusing language for all the parsers. We also inspected some AMR parsers and evaluated HanLP (He and Choi, 2021) with our dataset. The result shows the parser has strong bias finding the token that the quantifier phrase is modifying, and that the parser has some ability to distinguish sentences with syntactic ambiguity in other conditions.

Supervisor(s)

Daniel Hershcovich

External examiner(s)

Zeljko Agic

Time and place

14.06.2022 10:00

Online

Study Programme

Computer Science

Title

Uncertainty and exaggerations of scientific findings in social media

Abstract

Misinformation is a major element of today’s news cycle and social media. Therefore, we examine exaggeration and uncertainty when it comes to social media data and press releases that are related to scientific studies. We also study the tools that can be used for this, which is MT-PET model and a certainty estimator. We create an extended MT-PET model to handle an uncertainty task. Even though we are able to find relevant data for our model to use, we hardly see any datasets that are appropriately labelled that we can use. To solve this problem we use active learning to further improve upon the viable data samples and we also improve the models’ verbalizers using PETAL. We demonstrate with this that the extended MT-PET model is capable of capturing elements of both uncertainty and exaggeration. By analyzing the collected social media data, we observe that some correlation between high certainty and exaggeration exists. Finally we are also able to observe that exaggerated claims on social media tends to have a higher traction.

Supervisor(s)

Isabelle Augenstein

External examiner(s)

Zeljko Agic

Time and place

16 June, 14:40-16:30

Room UP1-1-1-N116B, Universitetsparken 1, 2100 CPH Ø

Study Programme

Mathematics & Computer Science

Title

Analysis of Liquidity Provision within Uniswap V2

Abstract

This project comprises of an introduction to Decentralized Finance, its infrastructure, and Decentralized Exchanges, as well as an analysis of the various aspects that impact gains for liquidity providers of the largest Constant Product Market Makers: Uniswap-V2. A digital twin of the CPMM has been built in order to undertake simulations and analysis. Data collected directly from Uniswap has been used to set the initial parameters of simulations to recreate trading trends that are currently occurring in the market. Experiments are conducted varying the number of liquidity providers interacting, their initial holdings and tokens tradable. Moreover, the impact of different trading fees on liquidity providers’ returns has been investigated. Results show that providing strategically liquidity to pools proves to be a profitable strategy that yields positive annual returns up to 1%. Experiments indicate also that the presence of other active providers influences returns to different extents and in different way. Smaller liquidity providers are seen to benefit from the presence of wealthier providers, while bigger liquidity providers are negatively impacted by active providers with similar portfolios. As expected, gains are directly correlated to pools’ trading volumes, as they are the source of profit for liquidity providers. This can be seen by the strategies adopted by providers.

Supervisor(s)

Omri Ross

External examiner(s)

Patrick Bahr

Time and place

17.06.2022 11:00

UP1-2-0-04

Study Programme

Computer Science

Title

Efficient generation of map graphs

Abstract

This thesis assumes a basic level of knowledge in graph theory and optionally GIS (Geographic Information System) as most of the project will be working with map data.

The project will be working with real world data typically used in GIS software, where the goal is to generate networks based on the paths and their associated data (such as speed limits, capacity, length etc.). QGIS is open-source project, which works with spatial data visualization and processing. I will be drawing comparisons to one of the processing tools they provide to generate the shortest path between two points by creating a network from the paths. The main problem is the conversion of the data to a obtain a reasonable network, that operates as expected in most if not all cases, where QGIS falls short in some data-sets and I will cover these cases in this project and which improvements are to be made.

The methods I will be using to generate such a network, are range searching datastructures as it can significantly reduce the problem complexity to only look at neighbourhoods of paths. The reason for looking at a neighbourhood of paths, is that it is only expected to connect to other proximate paths, that could be considered to intersect.

The solution should also strife towards creating a network, s.t. it can solve general graph problems, such as, but not limited to, shortest path or fastest route (ETA - Estimated Time of Arrival). The solution will also be implemented as a prototype as means to test practical differences between QGIS’s solution for creating networks and my own.

The project did succeed at providing a better solution for generating networks than QGIS for the data-sets chosen. It provided better practical run-times in the prototype, but not theoretical as there might be hidden costs in the library used to assist with the porting of QGIS’s solution to the prototype. The proposed solution also shows better quality (connections between paths, where it is expected) and provided the functionality for shortest path, fastest route and more.

Supervisor(s)

Pawel Winter

External examiner(s)

Jesper Larsen

Time and place

17.06.2022 10:00

A102 på HCØ, Universitetsparken 5

Study Programme

Computer Science

Title

Detecting the different types of relationships between lifestyle factors and diseases with deep learning

Abstract

Background: Many natural language processing techniques have been widely applied in the biomedical area to explore the associations between genes and diseases. Lifestyle factors are also receiving increasing attention from the scientific community in the past few years, however, there is still a lack of systematic research to discover their connection with diseases.
Aim of the project: This project aims to create a deep-learning-based model that will be able to detect and distinguish between the various types of associations that lifestyle factors can have with diseases. It will then be possible to apply it to the entire scientific literature to predict possible pairs of lifestyle factors and diseases and to examine the nature of their associations.
Results: 400 abstracts were sampled from the biomedical literature and the relationships between diseases and lifestyle factors were annotated in those. In total 2031 relationships were annotated belonging to eight different types. This dataset of associations was used to fine-tune a BioBERT model. The average F-score of the best-trained model for the relation extraction task
was 60.86%.
Conclusions: In this project a corpus with high-quality lifestyle factors - diseases associations annotations was created for the first time. This corpus was in turn used to fine-tune a deep learning model that can detect different types of associations between lifestyle factors and diseases. Due to the complexity of detecting several types of associations, more high-quality annotated data are needed to develop better models.

Supervisor(s)

Anders Krogh, Lars Juhl Jensen, Katerina Nastou and Esmaeil Nourani

External examiner(s)

Federico De Masi

Time and place

17.06.2022 13:00

Panum (Room: tbd)

Study Programme

Computer Science

Title

Measuring Convexity in Polygons

Abstract

Computational geometry has long been an established field of computer science with applications in many different areas. Often the problems have both numerical and combinatorial solutions each with their own advantages and drawbacks.

A problem in this category is finding a measure for convexity in polygons. This is a challenge on many levels, as the definition of the problem can take many forms. The measurement can by be defined by the ratio of reflex vertices, the largest convex polygon contained, the probability two points the polygon can see each other or any other quantity that is bounded by one.

We focus on the probability of two uniformly sampled points in the polygon being able to see each other, known as the Beer index. By definition this measurement lend it self well to approximation by numerical methods but different combinatorial solutions also exist to the problem. We give an overview of different sampling based algorithms to approximate the index, and the different combinatorial solutions, each with their own tradeoffs. These are weighed against each other for implementation purposes and a select few are implemented.

We explain the implementation details of the various methods, and issues faced when going from theory to practice are discussed. The performance of the implementations in practice are weighed against their expected performance, given by theoretical bound on the sampling algorithms.

Supervisor(s)

Mikkel Abrahamsen

External examiner(s)

Rüdiger Riko Jacob

Time and place

20.06.2022 11:00

Øvelseslokale 1-0-34 på DIKU, Universitetsparken 1

Study Programme

Computer Science

Title

Reverse mode automatic differentiation of\\histograms in Futhark

Abstract

This thesis presents a method of differentiating the parallel Futhark-construct reduce_by_index by integrating a new rewrite rule into the pipeline of the compiler, transforming the original code into a program that computes the adjoint of its input using reverse mode automatic differentiation. This thesis provides the relevant background of automatic differentiation and its forward- and reverse mode, along with the Futhark language itself. It will cover the main rewrite rule for reverse mode, and how it is used to derive the reverse mode rewrite rule for reduce_by_index. Lastly we cover an implementation of the derived rule, and evaluate its correctness and performance. 

Supervisor(s)

Cosmin Eugen Oancea

External examiner(s)

Mads Rosendahl

Time and place

Time TBD

PLTC meeting room, HCØ, Universitetsparken 5, 2100 CPH Ø

Study Programme

Computer Science

Title

Application-level data integrity enforcement in event-driven microservices

Abstract

Microservices are a new architectural approach for creating software systems as a
collection of self-contained, loosely connected modules with their data stores. Although separate services are designed and maintained independently to offer particular system features, such as data correlation, they frequently rely on one another. Asynchronous events combined with queue messages are a popular inter-service communication strategy. Event-driven microservices produce and broadcast events to other parties that need operating over such data or reacting to such external changes. Decentralized data ownership across microservices as a consequence has changed the response. Strong consistency guarantees are typically replaced by the eventual consistency model to protect the flexibility and availability of microservice systems. Because application developers must integrate a considerable amount of state management logic and validations at the application layer, the process brings additional obstacles. Application-level logic makes it harder to reason about the consistency of the application’s state management and its influence on the entire state of the application. To address these issues, we begin by examining prominent open-source microservice repositories for common patterns and deficiencies in application data management that result in anomalies. Based on discovered trends in the applications, we conducted a benchmark measurement of their performance and suggest three kinds of data constraints that developers may use to reason about data integrity. We apply constraint enforcement at the application level and investigate the implications on system performance. The imposed application evaluation demonstrates that adopting the restrictions eliminates the program’s data integrities at an acceptable cost to overall application performance.

Supervisor(s)

Prof. Yongluan Zhou and Rodrigo Laigner

External examiner(s)

Philippe Bonnet

Time and place

21.06.2022 10:00 - 11:00

Study Programme

Computer Science

Title

Reverse Automatic Differentiation in Futhark

Abstract

In this thesis, we describe our work with reverse automatic differentiation (AD) in the dataparallel programming language Futhark. The main motivation is to extend and optimise reverse AD in Futhark’s compiler to permit more expressive programs to by differentiated automatically and efficiently.

Futhark’s AD consists of a set of rewrite rules that are used to transform a program to its differentiated counterpart. We present reverse mode AD rewrite rules for the operations reduce-by-index and scan. Reduce-by-index, also known as multi-reduce, has a generic cases and multiple
special cases, of which the latter are loosely described by Schenck et al. 2022. We formulate and present a rewrite rule for the generic case and present specific rewrite rules for the special
cases as Futhark pseudo-code.

Likewise, we examine the reverse AD rewrite rules for scan presented by Schenck et al. 2022, one of which we have simplified with a performance benefit. The existing AD implementation is modified to work when the scan operates on tuples. We have extended the generic case with specialised rewrite rules for scan operators whose Jacobian matches specific patterns.
We have implemented the presented rewrite rules of both reduce-by-index and scan in Futhark’s compiler. The performance of differentiated programs is evaluated experimentally and compared to its primal program performance and the program differentiated with forward AD instead. In many case, this demonstrates reasonable reverse AD overheads and competitive performance to Futhark’s established forward AD implementation. 

Supervisor

Cosmin Eugen Oancea 

External examiner

Mads Rosendahl 

Time and place

Time TBD

PLTC meeting room, HCØ, Universitetsparken 5 

Study Programme

Computer Science

Title

Texture-based analysis of hippocampus for the prediction of Alzheimer's Disease using 2D Convolutional Neural Networks

Abstract

Alzheimer’s Dementia (AD) is a progressive and neurodegenerative disease, and as of today, it doesn’t have a cure. However, an early diagnosis of the disease can accelerate drug trials and improve the patients’ life. Brain changes start many years before the onset of symptoms, especially in the hippocampal area. These changes can be studied through MRI scans, and therefore, many biomarkers have been established in this area for the diagnosis of AD: volume, shape and texture. All of these biomarkers have successfully been studied using traditional Machine Learning (ML) tools that rely on manual feature engineering. In the case of texture analysis, a filter bank needs to be defined to extract the statistical patterns in the hippocampus.

In this study, we perform a texture-based analysis of the hippocampus using 2D Convolutional Neural Networks (CNNs) using T1-weighted MRI scans from the ADNI database. CNNs learn filters of increasing level of abstraction to get a set of features from each input, based on these features, binary classification is performed. Using off-the-shelf networks
(ResNet50, EfficientNetB0 and MobileNetV2) pre-trained on ImageNet, we build AD classifiers for different orientations of the bilateral hippocampus and fuse their scores. To preprocess the data and segment the hippocampus we use two different pipelines: FreeSurfer and FAST-AID Brain. To alleviate the problem of data scarcity we use data augmentation
techniques.

The results show benefits in using pre-trained networks and the importance of proper regularisation when training Deep Learning (DL) models using small datasets. Lightweight models with smaller capacity (EfficientNetB0 and MobileNetV2 networks) outperformed those with higher complexity (ResNet50). Furthermore, data augmentation gave consistently better results for the lightweight networks. We also showed the importance of image registration to help convergence and generalisation.

Supervisor(s)

Mads Nielsen and Mostafa Mehdipour Ghazi 

External examiner

Rasmus Reinhold Paulsen 

Time and place

10:00

Vibenhuset, Lyngbyvej 2, 2100 København Ø

Study Programme

Computer Science

Title

Sketching and Privacy for Anonymised Histograms

Abstract
In the realm of big data many statistical aggregates can be infeasible to compute due to space constraints. One approach to mitigate this problem is to work in a data streaming model where items of the large dataset arrives in a sequential manner. Algorithms working in this paradigm produces a compressed view of the dataset, often called a sketch, and
this can be used to approximate the wanted aggregates.

Such a data stream can be seen a multiset of tokens, meaning each token has a count on how many times it occurs. The aggregate of interest is the anonymised histogram, which is a multiset of the frequencies occurring in the data stream. In this thesis we will propose an algorithm to solve the problem of estimating the anonymised histogram.
The algorithm consists of a bottom-k sketch attached with counters and an unbiased estimator that provides an estimate on the number of distinct tokens that has frequency i in a stream. The estimator used by the algorithm will be analysed using two different approaches. In both cases we provide probabilistic bounds of the estimator to be within
multiplicative factors of the target in terms of an error parameter ε, and the results will be compared. In addition, a lower bound on the space usage will be shown, which shows the algorithm is optimal in some settings. The algorithm has been implemented and experiments has been made to showcase its performance in various settings and to
validate some theoretical findings.
The anonymised histogram also appears in sensitive contexts such as the degree distribution of social networks or password-frequency lists, which motivates the need of making the sketch differentially private. To analyse this aspect, we also introduce differential privacy for anonymised histograms and how a sketch can be make differentially private. Then we propose a potential method on how to estimate the histogram given a private sketch.

Supervisor

Rasmus Pagh

External examiner

Martin Aumüller

Time and place

10:45 at øv-3-0-25, DIKU (UP1)

Study Programme

Computer Science

Title

Simultaneous Synthesis and Segmentation of Brain MRI Modalities

Abstract
Automatic segmentation of brain tumors such as low-grade and high-grade gliomas from magnetic resonance imaging (MRIs) requires acquisition of multiple sequences, which include T1-weighted, post-contrast T1-weighted and fluid attenuated inversion recovery (FLAIR). However, scans acquired in hospital settings most often do not contain all modalities due to practical and logistical reasons, and as a result limit the quality of the obtained segmentation maps. Given the presence of limited data, the existing imputation methods for generation of missing modalities treat the imputation as a preprocessing step, and separate the process of synthesis from the process of segmentation. In this paper, we propose a method for simultaneous, end-to-end optimization of the synthesis network and the segmentation network, which produces realistic post-contrast T1-weighted images from T1-weighted and FLAIR images, and improves the segmentation of brain tumors from T1-weighted and FLAIR images only, with a mean per-class dice increase of 4.8%.

Keywords: Gliomas, MRI Modalities, Deep Learning, Brain Tumor
Segmentation, Simultaneous Synthesis and Segmentation

Supervisor

Mads Nielsen

External examiner

Rasmus Reinhold Paulsen

Time and place

11:00 at Vibenhuset, Lyngbyvej 2 

Study Programme

Computer Science

Title

Short-term prediction in patient-centered care

Abstract

Background: Healthcare is one of the most data intensive industries in the world, and with the field of AI, there is great potential to use data to move from prevention to prediction. Furthermore, AI tools are used for self-tracking technology which people are generating data from. Through these data we are enabling a more patient centered self-care environment. This is not only empowering the patient but also makes them more engaged with their disease.
Objective: This thesis attempts to examine how a possible temporary risk assessment may affect patients with an ICD. Furthermore, to find out whether a temporary risk assessment, can be considered valuable enough for the patient before the doctor has validated it. Also, we will be investigating their confidence in such technologies.
Methods: During the exploratory study we interviewed 4 patients with an ICD (Implantable Cardioverter Defibrillator) implanted, of 2 interview rounds. The aim of the first interview is to create an understanding within the patients' knowledge about their disease, AI, and the data they produce. The aim of the second round of interviews was to clarify the content preferences of patients between two different mock-ups. Here we used different qualitative methodological tools to collect and analyse the empirical data.
Conclusion: The analysis revealed a large heterogeneity among the four patients. It has been shown how AI can be used to make many tasks in healthcare more efficient. It also highlighted how some patient-centred care is neglected by AI-based tools. Validation from a doctor was presented as being decisive if disagreement between doctor and AI arose. It can therefore be concluded that a temporally risk assessment is not considered to have high enough value to stand alone before the physician has reviewed the algorithm's assessment

Supervisor(s)

Stina Matthiesen

External examiner(s)

Keld Bødtker

Time and place

21.06.2022 14:10 - 16:10

Sigurdsgade 41, 2200 Kbh N

Study Programme

Statistics

Title

Generalising Stein Variational Gradient Descent Through Rényi's alpha-divergence

Abstract

Stein Variational Gradient Descent (Liu and Wang, 2016) is a particle transport algorithm based on the KL divergence between two densities. We propose a wider class of algorithms based on Rényi’s α-divergence, Rényi’s α-divergence being a set of divergences that include KL divergence. We go on to show that there exist cases for which SVGD results in higher KL divergence between the target density and a kernel density estimation than an algorithm with similar parameters based on Rényi’s α-divergence.

Supervisor

Thomas Wim Hamelryck

External examiner

Lars Nørvang Andersen

Time and place

23.06.2022 13:00

TBD

Study Programme

Computer Science

Title

Futhark language server and VS Code extension development

Abstract
The thesis describes the implementation of a language server and VS Code extension for the Futhark programming language. Futhark is a young programming language and there are few tools that help developers code. So we developed a language server and a ready-to-use VS code extension that provides basic language support. We also discussed the workaround for error recovery, and some design choices of a language server. The Futhark language server is published as part of Futhark since its 0.21.9 release, and the VS Code extension is available in the marketplace with around 50 downloads.

Supervisor

Troels Henriksen

External examiner

Magnus Madsen

Time and place

TBD

Study Programme

Computer Science

Title

Embedding programmable touch sensations in digital text

Abstract
Reading is an inevitable part of our everyday tasks. Whether is in a computer, on e-book or on our mobile devices, we are constantly reading unconsciously. But, is there more information we can obtain from just plain text? Illustrations that accompany text have been doing that for decades, representing part of the information as an image. However, there might be different ways of retrieving information for a text than with the use of sight. What if we can incorporate another sense to enrich this reading experience? This report presents a software toolkit that provides
tactile stimuli whilst reading. These stimuli are called haptic effects and represent meta-information about the text that is being read. We conducted an exploratory user study where participants tested our implementation of the first version of the software toolkit that incorporated haptic sensations into text. Our results show that,
although participants agreed this technology might not fit light reading on adults, it has big potential for educational purposes.

Supervisor

Hasti Seifi

External examiner

Timothy Robert Merritt

Time and place

23.06.2022 16:40-17:30

Sigurdsgade 41

Study Programme

Computer Science

Title

Continuous Collision Detection in Softbody Simulators

Abstract

The department of computer science at the University of Copenhagen have an an ongoing physics simulator project in the works named LibRainbow. The simulator is written in Python and is intended to be able to simulate 3D scenes of both rigid bodies and soft-bodies. The simulator uses discrete collision detection, which is efficient but has multiple shortcomings with regard to accuracy. The goal of this project is to alter the functionality of the softbody portion of the LibRainbow simulator to use continuous collision detection, in order to avoid tunnelling. In this project I used Pybind11 to make a wrapper to a C++ library called Tight-Inclusion. This allows for very efficient estimates for the time of impact for collisions in the simulator. The bounding volumes k-DOP used in the narrowing phase is extended to include the candidate location. The contact determination is changed to use the current collisions instead of using penetrations. All of this together allows LibRainbow to run using a continuous collision detection scheme, though it has to be stated that the simulator still runs excessively slowly and there is a lot of optimization work still left to do.

Supervisor(s)

Kenny Erleben

External examiner(s)

Sarah Maria Niebe Abel

Time and place

24.06.2022 13:00

DIKU, Kenny Erlebens office, Univetsitetsparken 1

Study Programme

Computer Science

Title

Extension of Declarative Choreographies and Its Application in Distributed Systems

Abstract

Choreography programming has been proposed as a good way to design safe distributed systems. DCR choreography, based on DCR graphs notation, is a declarative model. The corresponding safety of DCR graphs is the basis for guarantee the safety property of a DCR choreography and the criteria of projectability also depend on the relationships in the DCR graphs. We extend DCR choreographies to include time and data constraints. We also implement tools to handle the extended DCR choreography, from designing to distributing the end-points in a real distributed system. To bridge the gap between theory and practice, we use conformance check to determine the difference between theoretical and practical semantics. We design algorithms to coordinate end-points in asynchronous systems and make the composition of end-points isomorphic to the choreography.

Keywords: DCR Choreography, Liveness, Conformance Check, Distributed Systems

Supervisor(s)

Thomas Troels Hildebrandt and Hugo Andrés López

External examiner(s)

Alceste Scalas

Time and place

27.06.2022 - 13:30

room 2.08 at Sigurdsgade 41.

Study Programme

Computer Science

Title

Phylogenetic Dimensionality Reduction in Non-linear Geometries

Abstract

This project analyses how dimensionality reduction can be performed on highly covariate biological traits. The geometry of the spaces in which these traits reside are often highly non-linear and thus a more novel approach than classical dimensionality reduction algorithms is required. The goal of these algorithms is to minimize the distorting effects of the implicit linear approximations performed on the surface of curved geometries. We first investigate the deep theory connected to both non-linear geometries, as well as the continuous trait models that model evolution as a stochastic process. Following this theory, we then propose two new algorithms; The first is a simple adaptation of a previously know algorithm of Polly et al.
(2013) to curved domains. The second is similar to ordinary PCA except that it considers small increments of a possibly non-linear brownian motion process, whilst implicitly removing the effects of biological covariance. We apply both our two new methods, as well as previously known methods to a range of synthetic datasets for visual comparison. Finally we apply our method to a dataset of traits from birds, AVONET (Tobias et al., 2022), and conclude that it does provide a similar estimate to the algorithm of Polly et al. (2013), but has additional flexibility built in. 

Supervisor(s)

Stefan Horst Sommer

External examiner(s)

Jes Frellsen

Time and place

27.06.2022

Small UP1

Study Programme

Computer Science

Title

Applications of Transformer Networks in Reinforcement Learning

Abstract

Increasing usage of Transformer networks has in recent years shown great
performance improvements of state-of-the-art models within several Machine Learning subfields, as they are able to encode long-term dependencies. Tangentially, Reinforcement Learning (RL) research has seen moves toward environments requiring the incorporation of long-term memory for agents to excel, due to the lack of Markovian rewards and state transitions. Such environments have commonly been addressed with simpler recurrent architectures, whilst applications of Transformers remain relatively understudied. We seek to study the applicability of Transformer networks within RL and to what extent they can encode long-term dependencies to improve performance in these environments. As such, we introduce common baseline models with and without memory mechanisms, and different Transformer variants. These models are evaluated on reactive RL tasks, short-term memory versions of classic tasks, and the Numpad environment as the primary task requiring long-term memory.

We show that models with simple memory mechanisms are able to perform well in both reactive, short- and long-term memory environments. Secondly, we show that Transformer-based models can be difficult to implement correctly and fine-tune for optimal performance. We conclude, based on our experiments in the RL setting, that other memory mechanisms remain superior in terms of learning speed and performance, given the same amount of training time. We further discuss improvements to our implementations, with the potential for decreased learning time and increased model performance

Supervisor(s)

Stefan Horst Sommer

External examiner(s)

Jan Frellsen

Time and place

27.06.2022 13:30

Small UP1

Study Programme

Computer Science

Title

Applying the method of Expander Decomposition to Fully-Dynamic Graph-Theoretic Problems

Abstract

We describe in great detail the (α, φ)-Expander Hierarchy described by Saranurak et.al. in [1] along with some of its applications. We also show that the Expander Hierarchy presents useful heurestics for the Vertex Labeled Distance Oracle problem which have potential for actual algorithms. Lastly, we use our gained understanding of the algorithm behind the Expander Hierarchy to obtain a framework that allows the user to solve certain problems which more quickly at the expense of an approximation factor.

Supervisor(s)

Jacob Holm

External examiner(s)

Eva Rotenberg

Time and place

TBD

Study Programme

Computer Science

Title

Reducing Synchronous GPU Memory Transfers

Abstract

We present a series of dataflow dependent program transformations that reduce memory transfers between a GPU and its host, and show how the problem of minimising memory transfers to the host amounts to finding minimum vertex cuts in a series of data dependency graphs. We provide a specialised algorithm to solve these minimisation problems, based on the Ford-Fulkerson max-flow algorithm, and detail techniques to model conditional execution and loops in a pure functional programming language.
We present our work in context of the array programming language Futhark, in whose compiler we have implemented our techniques. Empirical evaluation of 27 benchmark programs on four GPUs show mean speedups of 117–158%, heavily skewed by significant improvements to a few programs.

Supervisor

Troels Henriksen

External examiner

Patrick Bahr

Time and place

04.07.2022 14:00

PLTC-mødelokalet

Study Programme

Computer Science

Title

Total Resource Accounting in a Circular Economy

Abstract

In recent years, people are becoming increasingly aware of the issuesrelated to the way humans have treated the environment until now. More and more people want to do better, and so the demand for "green" products increases. But how can you guarantee that such resources are as environmentally conscious as they claim? With the help of software, digital twins of resources can be tracked and traced, and the supply chain modeled digitally, allowing for documentation of the complete history of resources, both natural and synthetic.

This report presents the design and implementation of an application that does just that. The application is based on existing models for algebraic resource accounting and discrete modelling methods. The application presented is a prototype, but does allow for modeling any supply chain in a circular economy.

Supervisor

Fritz Henglein

External examiner

TBD

Time and place

TBD