MSc Defences Winter 2021/2022

Study Programme

Nanoscience

Title

Generative Modelling of Sequential Data

Abstract

Autoregressive convolutional neural networks such as WaveNet are powerful models that have recently achieved state-of-the-art results on both text-to-speech tasks and language modelling. In spite of this, they have so-far been unable to generate coherent speech samples when learnt from audio alone. The original configuration of WaveNet uses repeated blocks of dilated convolutions to reach a receptive field of 300 ms. In this work, we test hypotheses relating to the role of WaveNet’s receptive field in learning to unconditionally generate coherent speech when not conditioned on auxiliary signals such as text. We also examine the usefulness of the learned representations for the downstream task of automatic speech recognition. By transforming the input data to stacks of
multiple audio samples per timestep, we increase the receptive field to up to 5 seconds. We find that enlarging the receptive field alone is insufficient to generate coherent samples. We also provide evidence that WaveNets create representations of speech that are helpful in downstream tasks. Finally, we find that WaveNets lack capability to model natural language and argue that this is the limiting factor for direct speech generation.

Supervisor(s)

Christian Igel

External examiner(s)

Peter Hagedorn

Time and place

9 February 2022 @ 09:00

Location TBD

Study Programme

IT and Cognition

Title

Preprocessing in Fair Binary-Classification

Abstract

Bias is a central problem in computer science, it is a multifaceted concept, but in recent years a great deal of research has focused on bias in relation to discriminatory and unfair outcomes. The use of biased datasets can e.g., lead to gender discrimination or ethnic discrimination, such discrimination
we believe is unfair. In this sense, bias mitigation then implies producing fair outcomes, albeit what is considered to be fair is up to debate. The field of fair classification is still developing its theoretical foundation by creating and comparing different fairness criteria; indeed, impossibility theorems
have shown that some fairness criteria cannot be mutually satisfied. This incompatibility of fairness criteria makes evaluating the fairness of a classification difficult, and it can also lead to disagreement between various stakeholders using conflicting fairness criteria. To remedy such conflicts a clearer understanding on how different fairness criteria and discriminatory bias mitigation methods affect each other on multiple datasets is needed.
The nascent field of fair classification has provided promising methods to mitigate bias in the form of discriminatory outcomes. Generally, such bias can be removed in one of three steps, preprocessing, in-processing and post-processing. Preprocessing, which is the focus of this thesis, is a methodological step that is applied before the main classifier of a machine learning pipeline. Preprocessing methods have shown themselves to be effective in mitigating discriminatory bias by relabelling, resampling, or
transforming the original data.
The main contribution of this thesis is that it examines and evaluates five standard preprocessing methods for the mitigation of discriminatory bias on four well-applied datasets in the literature. Traditional machine learning algorithms (logistic regression, support vector machines etc.) are
used on four binary classification tasks, further evaluation is made by comparing the effects of the preprocessing methods on three different fairness criteria. This cross-comparison between datasets, classifiers and fairness criteria can help researchers in their choice of classifier and fairness criteria, and it can also provide an understanding of the trade-offs between them.

Supervisor(s)

Maria Maistro

External examiner(s)

Thomas Riisgaard Hansen

Time and place

21.01.2022 10:00 - 12:00

UP1 1-0-14

Title

Reducing the Impact of Geo-Distribution Latency on Cloud-Based Mission-Critical Radio Communication

Abstract

A key requirement in every communication system is efficient and reliable call processing; a requirement which only increases in impact as the deployment of mission-critical radio communication networks begin to switch from on-site based servers to cloud services. Mission-critical
radio communication requires shorter response times and higher reliability than normal communication, and by deploying such networks into the cloud, they are put into a geo-distributed setting that on one hand can help reduce latency and improve system resilience and availability,
but on the other hand also requires data to be transferred between data centers. This might not be a problem for static or quasi-static data, but it can have a significant impact on the result when processing a mission-critical call which relies on dynamic data. Related data transfer protocols,
like ISSI for interconnecting radio systems or roaming protocols for wireless communication known from GSM, use an indirect approach where requests are channeled to a device’s home network regardless of which network it is connected to. The approach is robust and thoroughly
tested, but the approach requires multiple interactions between networks which takes time, and to live up to the requirements for mission-critical radio communication, we want to either have the data available or keep the transfer time as short as possible.
In this thesis we explore the potential of employing a direct approach for interconnecting networks in a fixed and distributed context by use of asynchronous data replication and call setup to reduce the impact of the transfer time in a geo-distributed setting. Through experiments
with a subset of Motorola Solution’s CirrusCentral Core deployed in two independent regions in Microsoft Azure, we show that the approach is feasible, but requires further exploration.

Supervisor(s)

Marcos António Vaz Salles (DIKU), Georg Lund Pedersen (Motorola Solutions)

External examiner(s)

Morten Frank

Time and place

20 January 2022 at 14:15

Online

Title

Hardness of Linkage Problems

Abstract

Polygonal linkages have long been studied, and this has lead to a large number theoretical questions concerning linkage reachability and reconfigurability, among other things. We give a survey of a selection of such linkage folding complexity problems. A number of these turn out to be PSPACEcomplete, so we also give a formal introduction so Turing machine and the PSPACE complexity. Much work has already been done by Demanaine et al., [DLO03], [CD17], to collect information on these subjects, but we go more in depth when in it comes to the PSPACE-proofs. Furthermore, we give our own small result, namely that reachability of a planar chain allowing for selfintersection is NP-hard, even with as few as 3 line segment obstacles. Thus we lower the lower bound of obstacles requiredx for NP-hardness by one.

Supervisor(s)

Mikkel Abrahamsen

External examiner(s)

Rolf Fagerberg

Time and place

Online

Title

Darkness in Context: Analyzing the Effects of Context on Dark Pattern Perception and User Actions

Abstract

Dark Patterns are elements of User Interfaces that trick users into doing something that they did not intend to do. While they have proven to be effective, the influence of context on Dark Patterns has yet to be seen. As context has the ability to alter choices, Dark Patterns might work differently on users depending on the context of a website. To analyze this, we developed two directly comparable interfaces with differing contexts, one commercial and the other non-commercial, and implemented Dark Patterns within. We conducted a user study in which participants tested one of the interfaces and analyzed both their performance regarding Dark Pattern susceptibility as well as their thoughts via interviews. Our results show that the context had little to no effect on how Dark Patterns affected participants, but had a noticeable influence on their thoughts and perception, with the main contributing factor being the trust of the entity behind the interface.

Supervisor(s)

Kasper Hornbæk, Irina Shklovski and Sonja Rebecca Rattay

External examiner(s)

Eve Hoggan Christensen

Time and place

14 January 2022, at 13:00

Online

Title

Development of a robust and reproducible preprocessing pipeline for Positron Emission Tomography (PET) data

Abstract

For several decades, neuroimaging techniques have been used to understand behaviour, memory and diseases such as Alzheimer's and depression. However, the neuroimaging community has recently started to acknowledge the existence of a reproducibility crisis, with the inability to reproduce results in an independent sample, and this has largely been
attributed to low statistical power, software errors, and flexibility in the data analysis. Positron Emision Tomography (PET) is an advanced neuroimaging technique used to quantify the concentration of molecular targets in the brain using intravenously injected radioactive tracers. The images acquired through PET undergo a series of data analysis steps (preprocessing pipeline) which often include motion correction, co-registration, delineation of volumes of interest (segmentation), partial volume correction and pharmacokinetic modelling. Previous studies indicate that the choice of preprocessing pipeline may significantly affect the outcome of study, and therefore there is an urgent need
to develop an open, robust and reproducible pipeline that can be used to process the PET data. In this project, we developed an open, robust and automated preprocessing pipeline for PET data using existing state-of-the-art neuroimaging software and implemented using Nipype. To test the pipeline for robustness and computational reproducibility, the pipeline was tested on three datasets from different scanners and tracers across different computational environments, namely a virtual machine, HPC server and a containerized environment using Docker. The resulting outcome measure from the PET data (the binding potential, BP) was found to be exactly the same using the containerized environment with Docker. However, we observed a significant difference in the BP estimates across subjects and brain regions across different computational
environments, indicating that different environments interact differently across subjects and brain regions to affect the final results. To further investigate this, we found that the differences observed across computational environments were attributed to differences in cortical and subcortical segmentations obtained by FreeFurfer when running
with different computational environments. The main outcome of this thesis is a a robust and reproducible PET preprocessing pipeline, openly available on GitHub. The developed pipeline ran successfully across different subjects using different scanners and radiotracers, and we therefore consider it to be robust and reproducible for future use.
However, testing on additional datasets is required to further validate the robustness of the pipeline, and new add-ons to the pipeline may be warranted in the future. 

Supervisor(s)

Melanie Ganz-Benjaminsen, DIKU and Martin Nørgaard, Center for Reproducible Neuroscience, Stanford University and Vincent Beliveau, Department of Neurology, Medical University of Innsbruck, Neurobiology Research Unit, Rigshospitalet

External examiner(s)

Kristoffer Hougaard Madsen

Time and place

14 January 2022, at 10:00

Online

Study Programme

Computer Science

Title

Detecting lung nodules from Computed Tomography scans

Abstract

Computed tomography (CT) screening has proven to be very beneficial in the early detection of lung cancer. With it, the number of CT scans to be examined also increases. That imposes a heavier workload on radiologists, who need to manually examine the scans to identify malignant lung nodules. Due to that, the interest in developing a computerized diagnosis has been increasing. There have been various attempts to find optimal methods and improve the accuracy of the machine-aided lung
cancer diagnosis. The use of Convolutional Neural Networks has demonstrated to give favorable results in many fields. The purpose of this thesis is to to experiment and implement CNNs for lung nodule classification by using 2D Convolutional Neural Network and 3D Convolutional Neural Network. The images are obtained from the
publicly available Lung Image Database Consortium (LIDC) dataset. The models aim to classify between cancerous and noncancerous nodules.

Keywords: lung cancer, lung nodules classification, convolutional neural networks

Supervisor(s)

Jens Petersen

External examiner(s)

Rasmus Reinhold Paulsen

Time and place

13-01-2022 13:00

Location TBD

Title

Measuring and mitigating gender bias in Vison and Language transformers

Abstract

We examine how the binary gender bias of bi-modal transformers are affected by the modal they are initialized from and how this bias can be measured and mitigated.

Supervisor(s)

Desmond Elliot

External examiner(s)

Zeljko Agic

Time and place

12 January 2022, at 10:00

Online

Title

Reducing the Need for General Anesthesia in Children Undergoing Neuroimaging by Preparation and Motion Correction

Abstract

Magnetic Resonance Imaging (MRI) is an important medical diagnostic tool. MRI suffers from long acquisition times on the scale of 30-60 minutes, which can lead to motion induced artefacts, especially in pediatric scans. As such, younger children are often scanned under General Anesthesia (GA) to prevent motion during image acquisition. An MR scan with GA requires additional personnel, occupies the scanner for, up to, six times as long as a regular scan and gives rise to concerns about possible health risks related to GA, such as cognitive neurodegeneration. Many methods to reduce GA in pediatric scans are being researched, such as preparation/training with a faux scanner and Motion Correction (MoCo).
Naturally one wishes to determine the effectiveness, of both preparation and MoCo.

The aim of this thesis is to evaluate the effect of preparation on anxiety levels in children, particularly if using a mobile phone application beforehand has an influence, and measure the quality of MR images using image quality metrics in an attempt to imitate observer scores as much as possible.

I will demonstrate that preparation/training positively affects children’s anxiety levels, as they drop significantly during the course of preparation. When modeling head displacement of the children inside the faux/mock scanner we found that age is a significant predictor indicating that older children move less. The age of the children and whether or not they used an app before the session proved to be significant predictors when modelling the time it took for a child to willingly enter the
mock scanner. Indicating that elder children and children who used a mobile phone app beforehand needs less time before entering the mock scanner.

Three mathematical based non-reference metrics, Average Edge Strength (AES), Co-occurrence Entropy (CoEnt) and Tenengrad (TG) were implemented and evaluated on more than 800 MR images, with 500 of the images scored with the golden standard, expert observer assessment. The Spearman correlation between metric scores and observer scores was calculated. This showed that AES had the highest correlation to observer scores and and in addition AES had clearer distinction between scores when MoCo was on compared to off. Thus it is recommended to use AES should one choose between AES, TG or CoEnt.

Supervisor(s)

Melanie Ganz-Benjaminsen

External examiner(s)

Kim Knudsen

Time and place

Hybrid

Study Programme

Computer Science

Title

Using Automatic differentiation to find gradients for recurrent neural networks in Futhark

Abstract

In this report, we will explain how Backwards automatic differentiation, can be used to transform an implementation of a recurrent neural network of the programming language, into a program that can find gradients for the weights that are given to the network. We will give the relevant background for neural networks, Futhark, and gradient descent as background. We will
present different methods for finding gradients, discuss their pros and cons and explain why we chose the method of backwards automatic differentiation. We will explain how the main rewrite rule of backwards automatic differentiation, can be used to create new rules for constructs in Futhark, and then we will use those rule to transform an implementation of the Elman neural network, into a program that finds gradients for the input weights. Lastly, we will validate the results against an implementation in the machine learning framework PyTorch, and discuss future work.

Supervisor(s)

Cosmin Eugen Oancea and Stefan Oehmcke

External examiner(s)

Mads Rosendahl

Time and place

TBD

Title

Computability in resource-bounded positional and other games

Abstract

Game theory is a branch of economics and mathematics that studies strategic interactions of completely rational, self-interested players who aim to maximize their payoff while taking into account the deliberations of the other players involved. In the infinitely repeated game setting, one way to calculate the payoff is by discontinuity. Namely, the payoff after
each repetition of the game is discounted by some fixed factor. Recent research provides necessary and sufficient conditions for all discount factors above some threshold, such that an infinitely repeated two-player normal-form game has a computable strategy with no computable best response. However, little is said about how a complete characterization
can be obtained for games in the extensive form. In this dissertation, we design a reasonable utility function that is used to derive a more generalized class of utility functions. This class of utility functions is used characterize a class of extensive-form games that satisfy certain property. Moreover, we provide a sufficient condition for all extensive-form games
to have computable strategies where no best response is computable.

Supervisor(s)

Jakob Grue Simonsen

External examiner(s)

Mads Rosendahl

Time and place

10 January at 13:40 - 14:40

Online

Study Programme

Computer Science

Title

Tracking And Diffusion Coefficient Estimation Of Simulated Quantum Dots In Simulated Two-Photon Microscopy Images

Abstract

The local transport of molecules can be understood by the underlying diffusion coefficient. Researchers have estimated diffusion coefficients using MSD analysis in widefield microscopy images [1, 2, 3]. However, widefield microscopy can not obtain deep tissue images. One way of obtaining deep tissue images is using Two Photon Microscopy (2PM) [4]. In this thesis we study if the MSD analysis commenly used for widefield miscroscopy images is also applicable to 2PM images.

MSD analysis requires trajectories. We use a KLT tracker to estimate trajectories from 2PM images [5]. A KLT tracker works by first detecting features in an image, and then track them using optical flow. The best object detection algorithms use deep learning [6, 7, 8]. Deep learning requires lots of training data. However, we do not know of any annotated 2PM image datasets.

To get annotated data, we construct a 2PM simulator. The simulator outputs simulated 2PM images and QD
trajectories. We have used the simulated data to finetune a pretrained Faster-RCNN object detection model. The finetuned Faster-RCNN model was implemented as the object detector for a KLT tracker [5]. Its performance was compared to KLT trackers using baseline object detectors (SIFT [9], Shi-Thomasi [10], and LoG [11]). We used the HOTA multiple object tracking evaluation metric to ensure that the evaluation of the trackers were not biased towards either detection or association [12].

Using the MS COCO standard [13], the Faster-RCNN model that was finetuned on 2PM simulated data, achieved a mAP[0.5 : 0.05 : 0.95] of 0.77, 0.75, and 0.66 on slow, medium, and high diffusion coefficient simulated datasets respectively. The best performing KLT tracker based on the HOTA score is the one using Faster-RCNN for object detection (KLT-FRCNN). KLT-FRCNN achieves a HOTA score of 0.59 for the dataset with the lowest diffusion coefficient. However, the HOTA score rapidly decreases as the diffusion coefficient increases.

Further, we show that our implementation of MSD analysis provides diffusion coefficient estimations that are correct on average. Both the 2PM simulated trajectories and KLT tracker trajectories are precise enough to correctly estimate diffusion coefficients with the MSD analysis on average. Finally, we use the finetuned Faster-RCNN model to detect QDs on real 2PM images. We find that the model is able to detect real QDs in real 2PM images, however with a lower mAP compared to its mAP on the simulated 2PM images.

We found that MSD analysis is applicable to simulated 2PM images. however, MSD analysis may not be applicable to real 2PM data because it contains fewer and shorter trajectories.

Supervisor(s)

Sune Darkner

External examiner(s)

Rasmus Reinhold Paulsen

Time and place

07-01-2022 - Online

Title

Design and Implementation of a Compiler for Fréchet

Abstract

Auto Differentiation (AD) is the pursuit of computing the derivative of a program. The has been the subject of research since the 1960s[1].

The Fréchet language is a domain specific language (DSL) in which programs are expressed in such a manner, that a special-purpose interpreter can derive expressions of the Fréchet derivative during interpretation. However, this approach produces expressions for the derivative that can be computationally demanding. Fortunately, this derivative is expressed entirely in terms of linear maps which are amenable to parallelization.

Futhark[2] is a data-parallel language with an optimizing compiler, that facilitates writing programs that can utilize massively-parallel hardware. It produces executables that can then be run on graphics processing units (GPUs).

In this thesis we investigate how we can speed-up computation of these Fréchet derivatives by utilizing Futhark. We design a DSL suitable for expressing Fréchet derivatives and implement a translator from this DSL to Futhark source-code. We further build an executor that compiles the generated Futhark source-code using the Futhark compiler and executes it on a GPU. Thereby enabling us to optimize programs and execute them on a GPU. We provide a test suite to justify the correctness of our implementation, as well as four micro-benchmarks and an example of a neural network written in the DSL. Finally, we visualize the performance characteristics over different input sizes.

Keywords: Auto Differentiation, Fréchet, Futhark, Linear Maps, GPU

Supervisor(s)

Fritz Henglein and Robert Schenck

External examiner(s)

Mads Rosendahl

Time and place

23 December 2021, at 14:00-16:00

Online

Title

Appearance models from 3D shapes and optical images for insects embedded in amber

Abstract

Image registration is the process of geometrically aling two or more images. In this thesis, we present an approach for building 2D to 3D image registration based on spline functions. The proposed algorithm minimizes the bending of paired images over bijective deformations. Thin Plate
Spline (TPS) interpolation is a commonly used 2D interpolation method.
It is physically meaningful in the sense that it assumes that there is a point
in the original shape, which corresponds to a new point under the new coordinates after the deformation.

Thin Plate Spline(TPS) interpolation is the numerical solution to this problem. Spline functions are used to map the 2D camera image to the flatten image, which can be obtain from libigl package in python. After that inverse the parametrization, map flatten image to the 3D model. This process completes the reconstruction from 2D to 3D. Place the camera in different positions to get multi-angle pictures. We can get the 3D model patches corresponding to the pictures. After maximizing the similarity of the images, they can be ”glued” together via transition and obtain a synthetic color appearance(RGB) of the 3D model.

Keywords: image registration, thin plate spline, parametrization, transition map, RGB

Supervisor(s)

Francois Bernard Lauze

External examiner(s)

Kim Knudsen

Date and Time

15 December 2021

Place TBD

Title

Encoding Representation for machine translation

Abstract

Algorithms that operate directly over meaning representations are becoming more popular in the field of Natural Language Processing. Most of current research has focused on developing new graph based algorithms for Natural Language Processing applications such as semantic role labeling, text classification and text generation. In this thesis I focus on expanding our understanding of the capabilities of this innovative family of algorithms.

To do so, in this thesis I apply Graph Convolutional Network (GCN) encoders to generate translation where the input is a semantic or a syntactic representation. While this has been done in the past, in
this project I aim to expand our knowledge of this model on two axes. First I use a challenge set approach to asses whether or not graph-based algorithms perform better than transformer machine translation systems in the context of Long Distance Dependencies. Then I dive into the application of graph convolutional network on low-resource machine translation.

I show that GCN encoders are able to boost the performance of machine translation systems trained over relative small number of examples. Then I present results that show that Long Distance Dependencies are also a challenge for GCN based machine translation systems, where only in some
cases GCN models outperform transformer based systems on the challenge sets that I utilize.

Supervisor(s)

Daniel Hershcovich

External examiner(s)

Peter Dolog

Time and place

26 November 09:30

Online