Alzheimer’s disease (AD) is a neurodegenerative disorder prevalent in older adults, leading to loss in memory, cognitive, and executive function. A characteristic feature of AD is the accumulation of amyloid-beta (Aβ) plaques, which are extracellular deposits of Aβ protein primarily found in grey matter. These Aβ deposits can appear in different forms. The six primary Aβ deposits covered in this work are: diffuse plaques, cored plaques, compact plaques, coarse grained plaques, cerebral amyloid angiopathy (CAA), and subpial deposits. It is speculated that some of the plaques types may be more harmful than others. Given that AD primarily affects an older population, the question arises of how individuals with AD differentiate from cognitively healthy centenarians (people over 100 years old). Consequently, this work attempts to classify and analyse different Aβ types present in donated brain tissue of cognitively healthy centenarians who escaped dementia and individuals diagnosed with AD. The main goal is to identify differences between these two cohorts. However, a challenge in this process is that the brain tissues contain many Aβ plaques, making manual identification difficult in terms of time and labor. To address this, a fine-tuned ResNet50 Aβ plaque classifier was developed in this research that was integrated into an Aβ detection pipeline capable of localising plaques. The model was initially pre-trained on the ImageNet dataset through contrastive learning, and subsequently fine-tuned using few-shot learning with a small number of annotated samples (315). The annotations included the six primary Aβ types whose structure are known and well-defined, and three other anomaly Aβ types that
served to filter out irregular plaques in the unlabeled Aβ data. After performing 5-fold cross-validation, the fine-tuned models demonstrated an average accuracy of 85.71% and precision of 89.47% on the primary types. The final classifier used on the unlabeled Aβ dataset was an ensemble model that incorporated majority voting, combining the predictions of the five models trained during cross-validation. Aβ loads were calculated for each primary Aβ type based on the classifier’s predictions. It was observed that across all considered primary Aβ types, the centenarians’ Aβ loads were statistically significantly lower compared to the AD cohort. The lower Aβ load in centenarians also held true across the frontal, temporal, parietal, and occipital cerebral regions for each primary Aβ type. To partially validate the
model’s performance, correlations were computed between the predicted Aβ loads of the primary Aβ types and neuropathological assessments collected from 75 centenarians. These assessments are common in related works and serve as a benchmark for the ensemble model. They include the Thal Aβ phase, which categorizes the distribution of general Aβ in the brain; the Thal CAA stage, which measures the severity of CAA; and CERAD NP scores, which evaluates the spread of neuritic plaques in the brain, which are a subset of cored plaques. Consequently, statistically significant positive correlations were revealed between: the Thal Aβ phase and the Aβ load of all primary types (r ranging 0.59-0.73); the Thal CAA stage and Aβ load of predicted CAA deposits (r = 0.66); and the CERAD NP scores and Aβ load of cored plaques (r = 0.68). Since the correlated types coincide with what the benchmark staging schemes measure, the model’s predictions seems to align with existing literature.

Graph Neural Networks are widely used as useful tools to investigate graphs because they can learn from the topological structure of graphs. In practical applications, the graph’s structure can change over time, have errors or be subject to adversarial attacks. These perturbations negatively impact the accuracy of the neural network. The theoretical stability of graph neural networks has been analysed already and in this paper, the stability of graph neural networks is investigated experimentally. The performance of different perturbation strategies is compared to see how different perturbations impact stability.

GNNs are a powerful tool for learning tasks on data with a graph structure. However, the topology of the graph in which GNNs are trained is often subject to change due to random, external perturbations. This research investigates the relationship between 5 topological properties of graphs (assortativity, density, edge connectivity, closeness centrality, diameter) and how stable GNNs trained on graphs with different topological properties are against different perturbations. The analysis is conducted by first synthetically generating graphs with different topological properties and training a GNN using the generated graphs. The synthetic graphs are then perturbed, and the relative change in the GNNs' output is measured. These results are further supported by conducting the same process on three popular GNN datasets: Cora, CiteSeer and PubMed citations. Finally, relationships between the graph properties under investigation and GNN stability are inferred using the results obtained from both synthetic and real-world datasets.

Graph Neural Networks (GNNs) have emerged as a powerful tool for learning from relational data. The real-world graphs such models are trained on are susceptible to changes in their topology. A growing body of work in the field of GNNs' stability to topology perturbations is trying to characterise how models respond to those changes, providing valuable insights that have enhanced the robustness of GNNs to adversarial attacks. The past work in this field, however, has only approached stability analysis using spectral graph theory, which is not applicable to all kinds of GNN models. In this paper, we aim to extend the past work in GNN stability by proposing an algorithm for analysing the stability of GNNs with model-agnostic GNN explainability tools instead of the mathematical framework of spectral graph theory. We demonstrate that the outputs of explainability tools can encode useful insights into the stability of GNNs and present a case study on using those insights to analyse node removal, edge removal, and edge weight perturbations.

Graph Neural Network holds significant impor- tance in various applications. Pioneering research has demonstrated state-of-the-art performance in practical applications such as Fraud Detection, Recommender Systems, or Traffic Forecasting by utilizing various Graph Neural Networks (GNNs) architectures. For these applications, one of the most important properties that needs to hold is the stability of GNN under stochastic perturbation as real-life networks undergo changes in topology on a frequent basis. However, it remains unclear how different architectures preserve this property under different perturbations. In this research, we aim to shed light on if this stability property undergoes drastic changes in the graph underlying topology, and if it affects the overall performance of the GNN in Traffic Forecasting problems. We demonstrate that the architectures differ in the stability property measured by different metrics, while some archi- tectures retains their state-of-the-art performance, providing useful insight on the analysis of stability property on different graph neural network archi- tectures in Traffic Forecasting problem.

As global urbanization reaches an all-time high, effective urban management becomes a crucial factor for efficient development. Enhanced monitoring of these transformations leads to more informed decision-making by policymakers, emphasizing the importance of tracking these changes. One method for monitoring is Change Detection (CD), which involves comparing two satellite images captured at different times to detect changes over a period of time. CD involves numerous difficulties, such as data collection, varying weather conditions, limited availability of datasets, noise, illumination differences, and discrepancies in the equipment used for image capture. Convolutional Neural Networks (CNNs) can address these issues by delivering more effective models with better performance than non-deep learning models. However, the rise of Transformers has led researchers to develop networks based on Transformer architecture, yielding more promising results than CNNs when more data is available. This paper conducts an analysis of two existing Transformer-based models, emphasizing the challenges of handling CD with artificially small datasets. Using smaller datasets reduces the requirements for remote sensing capabilities of satellites, simulating the limitations encountered during data collection and processing. The models under examination are the Bitemporal Image Transformer (BIT) and the Visual change Transformer (VcT).

Change detection with remote sensing data highlights se- mantic differences in an area between two or more time intervals. It involves the comparison of aerial photographs of the same location taken some time apart. This faci itates mass scale analysis of urban and rural data over time, including population trends, city expansion trends and illegal building detection. State-of-the-art methods for the task are predominantly deep learning networks, following an encoder-decoder architecture. These architectures all share the trait of having a ”fusion” point - a location in the network where inputs transition from being processed independently to becoming correlated. F sions can be classified in three categories: early, middle and late, depending on how deep within the network they occur. This study aims to show how changing the fusion impacts the size, spread and number of changes detected. It is motivated by how the receptive field of feature maps in convolutional neural networks expands in deeper layers, extracting features with different complexities. For this, four fusion architectures on three different datasets are compared: LEVIR-CD, HiUCD and a new, fully-controled dataset, CSCD. In terms of test accuracy and the changes’ size and spread, results are inconclusive. Which fusion achieves the highest performance varies per dataset. Possible reasons why include the complexity of remote sensing data and general differences between areas, but this is a subject of further study. The only conclusive category is the number of changes detected. On aver- age, all architectures overestimate the number of changes in a scene. When the accuracy of architectures is com- parable, however, early fusion overestimates the number of objects changed the most, while middle and late fusion give more realistic estimates. The case study has room for refinement in problem isolation, more data and extending the problem towards more architectures, but is a promising step towards understanding fusion.

Urban change detection involves identifying and analyzing alterations in urban landscapes over time. This process is crucial for urban planning, environmental monitoring, and disaster management, as it provides insights into urban growth, land use changes, and human impact on the environment. This study focuses on Recurrent Neural Networks (RNNs) due to their ability to capture temporal dependencies, making them suitable for analyzing changes over time. In this research, an RNN model, specifically the SiamCRNN, was trained and evaluated on multiple datasets representing different urban scenarios. The model performed well in detecting urban changes, especially in datasets with higher spatial resolutions, but faced challenges with datasets characterized by high spectral range and complex urban structures. These findings underscore the importance of spatial resolution in influencing RNN model effectiveness for urban change detection tasks.

The detection of changes in an area over time using remotely sensed data such as images is referred to as change detection. It has a large range of applications. For example, changes in buildings can analysed for urban planning. Many conventional image processing and machine learning-based algorithms have been developed for the purpose of change detection. Conventional non-classification algorithms have advantages in their reduced computational cost. Remotely sensed images vary in their spatial resolution, which is the area a pixel covers on the Earth surface. This work aims to explore how the spatial resolution impacts conventional non-classification pixel-based techniques in the urban change detection context, to provide insight into their performance with regards to detecting urban-related change over different resolutions. A systematic experiment is conducted by considering the LEVIR-CD and OSCD test sets in their initial as well as multiple downsampled resolutions. The change detection algorithms Change Vector Analysis (CVA) and Iteratively Reweighted Multivariate Alteration Detection (IR-MAD) are applied to the data individually. For creating binary change labels on a pixel-level, the Otsu algorithm is applied. A set of performance metrics is calculated, and trends in the metrics values over the resolutions are analysed. The data shows some trends towards improved metric values for lower spatial resolutions. The degree of the trends varies and is dependent on the algorithm and dataset. Overall, further research is necessary to consider influencing factors such as the amount of pixels in images, to refine the processing steps, and to broaden the scope of the experiment.

The border gateway protocol (BGP) is what holds the internet together by making data routing possible between various points on the internet. It is used to exchange routing information between and within networks on the internet with the use of special BGP routers. This routing information consists of a network path and its destination (IP prefix), and is stored in the routers' local routing tables.

Once a pair of BGP routers have been set up, they implicitly trust routes that are shared amongst them. This implicit trust system poses a problem called BGP hijacking: a malicious actor that takes over control of a BGP router, can simply start announcing IP prefixes they do not own. This can lead to different routing paths and thus redirected network traffic, possibly to the infrastructure of the attacker.

In 2017 there were a total of 13,935 routing incidents, consisting of both inadvertent misconfigurations and deliberate IP prefix hijacks. Over 10% (6,000) of all autonomous systems (AS) were affected and about 5% (3,106) of all AS's were a victim at least once. Despite numerous possibilities that do exist in terms of securing BGP, these numbers show that BGP hijacks are still prevalent. As such, there is a lot of room for improvement in terms of preventing hijacks and mitigating their impact.

The aim of this research project was to provide a proof-of-concept framework to aid in performing threat intelligence. A methodology was developed in which the hijacked IP prefixes from BGP hijacks are used as a foundation to gather information about what domains lie in the scope of the impact. Features were determined and extracted from the data, to say something about what domains are more likely to have been targeted. The methodology could also be used as a framework for a real-time detection and response system, by maintaining a live dataset of the different indicators.

Reports of BGP hijacks were analyzed to identify hijacked prefixes. These were referenced against DNS servers to determine what DNS traffic was redirected. The NS records of those servers were then used to identify what domain queries were redirected. Multiple indicators were devised to indicate if domains might have been redirected. This included SSL certificates, DNS records, IP prefix resolutions, as well as the hosting AS's.

We believe this framework serves as a tool to be used for threat intelligence. It uses the basic ideas of BGP hijack detection to look deeper into the impact of such incidents. Instead of only looking at the hijacked prefixes themselves, the DNS servers and domains that are within the scope of the hijack are also considered along with what indicators they might present. If a BGP hijack takes place, it often takes multiple hours to get resolved. Afterwards, the incident is analysed and it is determined what happened. One of the reasons for this long winded process is that only a relatively small part of the internet is usually affected, so it takes a while before things get noticed. This framework could also be used as a foundation for a detection and response system, such that domains that might be targeted and redirected during such a hijack can act in time.

One important limitation of this research is the incompleteness of the available data. This could be solved by using more data sources and tracking data over time. Another issue in a real-time detection and response system would be the sparse visibility of BGP sensors. Adding other sensor networks could increase the covered surface area of the internet. Each indicator of compromise could be refined further, to better indicate malicious activity.

Arguably the main goal of artificial intelligence is to create agents that can collaborate with humans to achieve a shared goal. It has been shown that agents that assume their partner to be optimal can converge to protocols that humans do not understand. Taking human suboptimality into consideration is imperative to perform well in a coordination task. One way to achieve this is imitation learning, where you train an agent on recorded data from a human playing optimally. I created several agents using different implementations of behavioral cloning, by reducing the dataset to state-action pairs and training a neural network on this. To evaluate their performance, I used an environment that poses a coordination challenge based on the popular game Overcooked. Neither expanding nor reducing the feature space that the agents are trained on yielded any significant improvement in the performance of the agents. In fact, expanding the feature space to include some historical data made the agent less generalizable and especially failed to perform when paired with agents with unfamiliar strategies. These limitations were mostly posed by the available dataset, which was not big enough to support more features and of too low quality of gameplay to create agents that perform exceptionally well.

Overcooked, an immersive multiplayer video game centered around cooperative cooking challenges, provides the roots for this research project. The study focuses on designing and evaluating a hand-authored controller in comparison to controllers implemented using various machine learning techniques, such as Population Based Training, in the context of a simplified version of the game. The main objective is to assess the cooperation between the hand-authored controller and a human controlled agent, with a particular emphasis on the coordination and minimizing errors during runs of fixed time length. A testing method has been designed in order to more accurately evaluate the performance. Through the implementation of a specialized controller utilizing techniques such as Behavior Trees and Decision Trees, the controller was able to successfully accomplish the task of delivering soups. An analysis was made to examine the performance of the hand-authored controller in comparison to the results obtained in other research papers that use the same simplified version of the game. After the results have been obtained, a clear difference was visible. The scripted AI performed better when paired with himself than one created with population based training that was paired with himself. However, the scripted AI was not better than a coupled planning algorithm that was paired with himself, but the computations were easier and faster for the scripted AI. When paired with a human, the overall performance decreases, but increases only for one specific map.

In ad-hoc cooperative environments, the usage of artificial intelligence to take supportive roles and work in collaboration with humans has proven to be of great benefit. The objective of this research is to evaluate the use of population-based training for reinforcement learning agents in a simplified version of the multiplayer game - Overcooked. The method used to answer that question involves evaluating the performance of the agents when paired with a human proxy and their learning curves on different layouts. Based on the employed method, it was concluded that both PBT and other self-play agents display notable underperformance when compared to human proxies and agents trained using human data. Moreover, while the inclusion of mutated agents enhanced sample efficiency in layouts with minimal collision risks, its effect on the final performance of PBT in those layouts was negligible. However, this approach managed to improve performance in layouts where collisions were the primary limiting factor.

The popular video game "Overcooked" is a great example of a task requiring complex planning and cooperation with other players. This game is used as the inspiration for an environment for evaluating AI, called "Overcooked-AI". This paper implements a centralized critic into the Overcooked-AI environment's implementation of the PPO algorithm and compares the results with the decentralized critic approach when it comes to cooperation with human-like agents and computational efficiency. \\

The centralized critic approach gives similar results compared to the decentralized critic approach, both in self-play and when playing with human-like agents. This is probably due to the decentralized critic approach already having full access to the entire observation space, and no hyperparameter tuning being done due to a lack of time.

Cooperative AI is AI designed to cooperate with humans. One example of such an AI, made using planning algorithms, was studied in a paper from 2019 which used a simplified version of the video game Overcooked for evaluation. However, only limited evaluations were possible due to the long runtime and heuristic optimisations made. This paper will attempt to increase the amount of evaluations while removing an important optimisation limiting the functionality of the AI: the omission of counters. In the end it will find ways of reducing the runtime and increasing performance, which together allow for the addition of counters in a few instances. The final results with counters suggest there is more to gain upon removal of a select few heuristic optimisations and improvement of the human behavioural clone used to simulate a human.

In real-life scenarios, there are many variations in sizes of objects of the same category and the objects are not always placed at a fixed distance from the camera. This results in objects taking up an arbitrary size of pixels in the image. Vanilla CNNs are by design only translation equivariant and thus have to learn separate filters for scaled variants of the same objects. Recently, scale-equivariant approaches have been developed that share features across a set of pre-determined fixed scales. We further refer to this set of scales as the internal scales. Existing work gives little information about how to best choose the internal scales when the underlying distribution of sizes, or scale distribution, in the dataset, is known. In this work, we develop a model of how the features at different internal scales are used for samples containing differently-sized objects. The proposed model return comparable internal scales to the best-performing internal scales for different data scale distribution of various width. However, in most cases, the scale distribution is not known. Compared to previous scale-equivariant methods, we do not treat the internal scales as a fixed set but directly optimise them with regard to the loss, removing the need for prior knowledge about the data scale distribution. We parameterise the internal scales by the smallest scale which we refer to as σbasis, and the Internal Scale Range (ISR) that models the ratio between the smallest and largest scale. By varying the ISR, we learn the range of the scales the model is equivariant to. We show that our method can learn the internal scales on various data scale distributions and can better adapt the internal scales than other parameterisations. Finally, we compare our scale learning approach and other parameterisations to current State-of-the-art scale-equivariant approaches on the MNIST-Scale dataset.

t-distributed stochastic neighbor embedding (t-SNE) is a dimensionality reduction technique able to embed high-dimensional data into 2 or 3 dimensions for the purpose of visualization, and has proven to be useful in various applications, such as single-cell analysis. Normally t-SNE embeds into Euclidean space, but recent work shows that embedding into hyperbolic space can lead to higher quality embeddings. Most notably on data containing hierarchical structures, as seen in single-cell measurements. As such it would be interesting to generalize computation of t-SNE embeddings to embeddings in hyperbolic space. A missing tool for this is the approximation of the objective and the gradient. This is needed as the cost for evaluating the objective and gradient is quadratic in the number of data points to be embedded. In practice such approximations are already used for Euclidean embeddings. In this thesis we introduce the first approximation scheme for hyperbolic t-SNE embeddings. To achieve this we generalize a Barnes-Hut approximation scheme to hyperbolic space. A major difficulty for this is to transfer the use of a quadtree in the Euclidean plane to hyperbolic space. We propose the usage of a polar quadtree in the Poincaré disk and show that our solution yields substantial gain in run time, in particular when embedding larger data sets, with the resulting embeddings being similar in quantitative and qualitative aspects.

Commonly, when researchers are figuring out the effect of a putative cause, additional variables influence the cause and the effect. These are called confounders, and they obfuscate causal relationships. Inverse Probability Weighting is a method that can be applied to remove confounding and show a causal effect. This study aims to determine if Dota 2 game outcomes can be predicted based on team composition diversity and if 'Inverse Probability Weighing is a helpful tool for this. First, metrics to assess team diversity were determined, and the two confounders, "Player skill" and "Hero skill", to the 'is-diverse' treatment and game outcome were identified. Next, a dataset with Dota 2 games was gathered containing all the relevant variables to measure the confounders. Finally, inverse probability weighting was applied to measure the Average Treatment Effect of the putative cause.
The results suggest that team diversity significantly impacts the game outcome. However, the results were close in most cases when comparing the results with data that was not corrected for by confounders. A possible explanation may be that the confounders didn't have enough influence on some diversity metrics since they were too vague. For example, the most complex diversity metric, which covered the most team-composition characteristics, showed a clear difference between the average treatment effect with inverse probability weighting being applied and not applied.

Dota 2 is one of the most popular MOBA (Multiplayer Online Battle Arena) games being played today. A Dota 2 match is played by two teams of 5 players. The main goal of the game is to destroy the opposing team’s Ancient tower, the team that manages to do so, wins the game. An essential part of a match is the hero selection phase before it starts. There are different ways to select heroes in different game modes, but the game mode that is used for this research is the Captain’s mode where each team is assigned a captain and the captains take turns picking and banning heroes. The main question that this research aims to answer is: “What is the effect of the Pudge hero being picked in a team on the outcome of a Dota 2 game?” Causal inference, a discipline that is concerned with discovering causal relationships using data analysis under certain assumptions about the data is used in this research to measure this effect. More specifically, the g-formula is the causal inference method of choice for this research. The data that is used for this research is gathered through the OpenDota API. After running the correctly formatted data through the g-formula implementation, the effect of the Pudge hero being picked in a team on the game outcome was estimated as -0.2848%. Meaning that, on average, there is a 0.2848% less chance of a team winning the game, if the Pudge hero is in that team.

The front-door adjustment is a causal inference method with which it is possible to determine the causal effect of applying a treatment given a setting which satisfies the front-door criterion. This involves having a mediator through which all the causal effect flows from treatment to outcome. The front-door adjustment adjusts for confounders and tries to only measure the causal effect from treatment to outcome. The goal is to test the applicability of the front-door adjustment using the game of Dota 2 as a testing ground. The front-door adjustment has been applied to find the effect of picking ‘Slark’ on the outcome of the game. The mediator in this case is the enemy team buying an item called ‘Hurricane Pike’. Two different approaches have been used, both giving varying results. These varying results lead to different possible interpretations. This variety of interpretations therefore suggest that the front-door adjustment is not a valid method for this specific scenario, likely due to the complexity of the game and perhaps the simplified representation of the game in the data-set.