Decentralized Learning is becoming increasingly popular due to its ability to protect user privacy and scale across large distributed systems. However, when clients leave the network, either by choice or due to failure, their past contributions remain in the model. This raises privacy concerns and may violate the right to be forgotten. In some applications, it is also undesirable to retain the outdated influence of clients that no longer reflect the current state of the system. While Machine Unlearning has seen significant progress in Federated Learning, similar solutions for Decentralized Learning are limited because there is no central server to orchestrate these operations. This work adapts and extends a state-of-the-art Federated Unlearning algorithm, QuickDrop, to operate in a decentralized setting. Our method uses synthetic data to reverse the influence of dropped clients and efficiently restore the model’s generalization performance. It also supports unannounced client crashes and performs reliably in sparse network topologies. We evaluate the algorithm on MNIST and CIFAR-10 using different graph structures and show that it remains competitive with oracle baselines that require access to sensitive data. Finally, we discuss the limitations of our approach and suggest directions for future work in Decentralized Unlearning.

Decentralized learning (DL) enables a set of nodes to train a model collaboratively without central coordination, offering benefits for privacy and scalability. However, DL struggles to train a high accuracy model when the data distribution is non-independent and identically distributed (non-IID) and when the communication topology is static. To address these issues, we propose DissDL, a fully decentralized topology optimization algorithm for DL where nodes select the peers they exchange models with based on local model dissimilarity. DissDL maintains a fixed in-degree while dynamically adapting the communication graph via gossip-based peer discovery and diversity-driven neighbor selection, improving robustness to data heterogeneity. Experiments on CIFAR-10 and FEMNIST show that DissDL achieves faster convergence, more stable learning, as quantified by lower inter-node variance across nodes, and higher final accuracy than with static topologies and state-of-the-art baselines. For example, on CIFAR-10, DissDL reaches the best accuracy achieved by its most competitive baseline using 1.34 × fewer communication rounds and 1.34 × less communication cost, while maintaining comparable overhead.

Decentralized Learning (DL) is a key tool for training machine learning models on sensitive, distributed data. However, peer-to-peer model exchange in DL systems exposes participants to privacy attacks. Existing defenses often degrade model utility, introduce communication overhead, or are not applicable to a DL system. Model chunking (splitting a model into parts before sharing them individually) has been proposed as an alternative, but its standalone privacy implications have not been investigated. This work implements and evaluates three distinct model chunking methods (static, cyclic, and random) against two privacy attacks: membership inference and linkability. Our work introduces a Hungarian matching-based enhancement to the linkability attack, and relaxes prior assumptions by evaluating attack success under limited access to neighbor datasets. Our results show that chunking increases vulnerability to membership inference. However, static and random chunking are effective against linkability attacks under specific conditions, particularly when full epochs are used during training.

Decentralized learning (DL) enables collaborative model training in a distributed fashion without a central server, increasing resilience but still remaining vulnerable to the same adversarial model attacks, that Federated Learning is vulnerable to. In this paper, we test two defense mechanisms, specifically designed to protect fully decentralized learning, in the scenario of chunked models and introduce one new defense. We clearly define the threat model, where malicious peers either try to perform a backdoor attack or an untargeted label flipping. The defenses are Norm Clipping, Sentinel and we propose Adaptive Norm Clipping. We evaluate the effectiveness of these defenses on the CIFAR-10 dataset. Our results indicate that the model attacks (backdoor attack and untargeted label flipping) significantly harm model accuracy on vanilla DL with chunking. While static defenses like Norm Clipping and Adaptive Norm Clipping reduce the impact of the attack, they lower the final test average accuracy in the chunked scenario. Also, robust aggregators like Sentinel fail at mitigating the attack, but do not lower the average test accuracy.

AlphaZero and its successors employ learned value and policy functions to enable more efficient and effective planning at deployment. A standard assumption is that the agent will be deployed in the same environment where these estimators were trained; changes to the environment would otherwise violate their expectations and could result in suboptimal decisions. In this work, we investigate how environment changes affect the usability of the learned estimators and develop criteria that can quickly detect and localize such changes. Moreover, we develop novel planning methods that leverage these principles as well as further modifications of standard Monte Carlo planning techniques. These methods demonstrate superior performance under several tested environment configurations. The main assumptions and limitations of our approaches are also discussed, providing a foundation for future research to broaden their applicability. The code is available on GitHub.

Complex reinforcement learning (RL) models that receive high rewards in their environments are often hard to understand. To this end, more interpretable models can be used, such as decision trees. To be able to deploy these models in safety-critical environments, they need to be high-performing and verifiable. Optimal decision trees can fulfill both these goal. While some methods already exist to find optimal decision trees, none have applied them to RL environments with continuous state
and action spaces [4; 13]. Broccoli is a state-of-the-art approach to synthesizing decision tree policies for blackbox environments [3]. Given a discretisation of the continuous state space, it can find the optimal tree in discrete action environments.
This research will focus on discretising action spaces of continuous environments. In doing so, discrete action spaces are created for which Broccoli can find optimal decision trees. Additionally, a goal is to show that it is possible to discretise continuous action spaces and compute corresponding optimal decision trees in feasible time. Furthermore, it proposes informed discretisation techniques that result in better-performing decision tree policies.

Reinforcement learning models are being utilised in a wide range of industries where even minor mistakes can have severe consequences. For safety reasons, it is important that a human expert can verify the decision-making process of a model. This is where interpretable reinforcement learning proves its importance. This research is focused on training decision tree policies with a limited size and evaluating them on continuous action space environments. For that, a DAGGER algorithm is used with appropriate modifications to account for the continuous setting. The results demonstrate that small decision trees can replicate the high-performing neural network policies (e.g., TD3), achieving close to benchmark scores. Therefore, it is possible to explain the complex model's behaviour with much more understandable structures.

Understanding the energy profile of a complex, multi-faceted software system is difficult. In this thesis, we present a novel methodology, called SMURF, a five-step methodology that gives insights into the energy consumption of a complex system. The methodology is broadly applicable, supports informed decision-making, and closely involves and engages stakeholders. We evaluate the methodology with a case study on MUST, a software system used in spacecraft operations. In the case study, SMURF successfully finds energy hotspots and wasteful components in MUST, and is used effectively to formulate actionable recommendations. Through the case study, we find that the SMURF methodology serves as an effective engagement tool to get developers, users, and product owners interested in sustainable software ideas.

Procedural Content Generation methods enable the creation of varied content algorithmically. Wave Function Collapse (WFC) is one such method. It is a tile-based local constraint solver commonly applied to world and map generation for grid-based content; it is able to create varied output from the same set of rules. While WFC is versatile, content generated with it is i) limited to one grid, ii) based on uniform tiles and iii) must use the same tile set and constraints throughout the grid. Due to these limitations, large classes of content are excluded, such as structured objects. Structured objects consist of an assembly of multiple components, each possibly based on a different tile set. We therefore propose Expressive Wave Function Collapse (XWFC), a major extension of WFC that enables solving and combining multiple grids with different Non-Uniform Tile (NUT) sets. Additionally, we can guarantee NUT shape and size preservation even under WFC’s Overlapping Model. With these generalizations, new domains are within reach for structured objects based on NUT sets, such as Tetris or LEGO.

Synthetic art is the end result of artificial intelligence models that have been trained to generate images from text prompts. "Comic synthesis" is one such use case, where comic illustrations are produced from textual descriptions. Previous attempts at comic synthesis have utilized conditional Generative Adversarial Networks (cGANs), but this approach has encountered challenges in generating consistent and visually appealing comic panels. Strict data requirements and quality limitations have left room for improvement. We propose a novel approach to comic synthesis using Stable Diffusion, a powerful generative modelling technique. The study investigates the fine-tuning of the stable diffusion model specifically for the generation of Dilbert Comics from textual prompts. We explore different techniques to fine-tune the stable diffusion model for comic synthesis including Dreambooth and LoRA. Through extensive experimentation and analysis, with an FID score of 123, results produced using the Lora technique outperformed Dreambooth, excelling in understanding the Dilbert style, while Dreambooth struggled with multiple-subject training. Results are also compared with previous approaches based on conditional GANs. While the quality and detail greatly improved, the transition from conditionals to text descriptions meant the results were less accurate. The results show the potential of stable diffusion in generating appealing Dilbert Comic panels while highlighting the need for further exploration to enhance the alignment between textual descriptions and the generated images.

Face clustering is a subfield of computer vision and pattern recognition with many applications such as face recognition and surveillance. Accurate clustering of faces can also help us to create labeled datasets. However, in the domain of comics, face clustering is not well studied. Therefore, it is uncertain which methods of feature extraction and clustering perform well on faces of comic characters. In this paper, we investigate the effectiveness of comic face clustering. To conduct our investigation, we implement two pipelines: one that automatically extracts character faces from comic strips, and another that clusters the extracted faces. Using Dilbert Comics for our experiments, we examine the performance of various feature extraction and clustering methods. Additionally, we experiment with combining feature extraction methods and removing noisy samples to increase the clustering accuracy. We show that using color information is crucial for accurate clustering, and combining color with shape features further improves accuracy. However, our experiments indicate that accuracy improvement is not guaranteed for every combination of feature extraction methods. We also demonstrate that removing noisy samples using hierarchical clustering can increase clustering precision. Using our findings, we achieve an F1 score of 0.752 based on our Dilbert Comics dataset of 77,768 face images. We obtain this result by clustering 20,988 non-noisy face images into 35 clusters with a precision of 0.886.

Contrastive Language-Image Pretraining (CLIP) has gained vast interest due to its impressive performance on a variety of computer vision tasks: image classification, image retrieval, action recognition, feature extraction, and more. The model learns to associate images with their descriptions, a powerful method which allows it to perform well on unseen domains. Often, the descriptions fail to capture text which is contained within the image, a source of information which could prove useful for a handful of computer vision tasks. This limitation requires finetuning in domains where contained text is important. In fact, CLIP has mixed performance on Optical Character Recognition (OCR). This paper proposes a novel architecture: OSBC (OCR Sentence BERT CLIP), which combines CLIP and a custom text extraction pipeline, composed of an OCR model, and a Natural Language Processing (NLP) model. OSBC uses the text contained within images as an additional feature when performing image classification and retrieval. We tested the model on multiple datasets for each task, occasionally outperforming CLIP when images contained text, while maintaining finetunability, and improving the model's robustness. In addition, OSBC was designed to be generalizable, meaning it is expected to perform well on unseen domains without finetuning, though this was not achieved in practice.

Cancer is one of the leading causes of death in the world. While there have been many studies investigating the development and progression of cancer in human tissues using host omics data or microbial data, there is a lack of research combining both types of data, even though both modalities have been shown to affect cancer morphology and aetiology. Studies which do combine these modalities often use simple methods or do not consider the relation between the two modalities and disease phenotypes. Such an integrated approach could offer additional insights and lead to the discovery of new disease biomarkers and better treatment strategies and therapies.

In this paper, we investigated whether such a holo-genomic approach offers additional information compared to using the modalities separately, by comparing the performances of prediction models built using the individual and integrated modalities for various prediction endpoints. To do this, we used TCGA gene expression data for the host omics modality and bacterial genus abundance data from the TCGA-mined Cancer Microbiome Atlas (TCMA) for the microbiome modality.

We found no improvement when integrating host gene expression with microbial abundance information compared to using the gene expression data individually, and the microbial data provided the least amount of diagnostic information. This is likely due to the information density of gene expression data, high variation of the microbiome, and the quantity, specificity and validation of the TCMA data. These results suggest that the holo-omics approach might not provide additional utility in certain contexts, that additional considerations have to be made when choosing microbial and host omic datasets for holo-omic integration, and provide an insight into the usability of the TCMA data set.

Visible Light Communication is a method of wireless communication that avoids the oversaturated frequencies used by radio communication. Prior research typically uses the camera on smartphone for communication, but using a camera is energy-intensive and inefficient. Some alternatives are photodiodes and ambient light sensors, however the former is often not available on phones and the latter is too slow to be useful. This thesis introduces a system, JAB, which combines two sensors to achieve high throughput while keeping energy usage low. The impact of distance, modulation frequency and light are measured on the system implemented on an Android smartphone. Overall, the system is usable and does greatly improve the energy effiency of using VLC. An important next step is creating a custom hardware setup with a kernel driver implementation for a full system test.

Visible Light Communication (VLC) is becoming an important research area where Visible Light Sources such as LED, Halogen Lamps and even the sun can be used for Wireless Communication. LED-to-Camera Communication is a form of VLC, where the camera can notice intermittent stimuli changes above a certain threshold frequency (flicker fusion frequency) which human eyes cannot no- tice. There exist numerous methods to encode data (= Modulation Schemes) for LED-to-Camera Communication, however, with some modulation schemes, VLC cannot support any practical appli- cations because the achievable data rate either is too low or the system is difficult to be implemented on off-the-shelf devices.
In this paper, we will propose a system design for LED-to-Camera Communication with Color Shift Keying (CSK) as a modulation scheme to encode the transmitted data. In Color Shift Keying, the sys- tem makes use of colors to encode bits, the receiver will then decode the received color into the corre- sponding bits. The proposed design provides a sys- tem outline which may obtain a more practical data rate by using CSK as a modulation method and a CMOS image sensor as receiver. A CMOS image sensor consists of a matrix of photodiodes which combined make up the captured image. The rolling shutter effect makes use of this matrix-like struc- ture by scanning each line separately, by using the rolling shutter effect the camera captures lines with different colors which can be decoded into bits.

Increased urbanisation has led to significant challenges for public transport operators. Inconsistent demand leads to peaks in passenger activity on the network. Moreover, the COVID-19 pandemic has introduced a need for social distancing as well, limiting the desired capacity of vehicles. To combat this, intelligent real-time and data-driven decision making is required. In many cases, the data required is lacking or not available in real-time. Our research addresses these challenges by providing means to gain insight into the passenger load of public transport vehicles. The focus of this research is to investigate how using in-vehicle sensor data can help in constructing an estimate of the passenger load and evaluate its contribution.

By combining in-vehicle sensor signals with historical passenger flow patterns, a novel fusion model
based on gradient boosting machines is constructed that can make real-time predictions of the passenger load using this data as input features. The evaluation shows that its estimates have a mean absolute error (MAE) score of 7.83, outperforming a random forest model baseline by 37%. Moreover, a crowding indicator analysis demonstrated that when predicting crowding indicators, the model achieves a weighted F1 score of 0.828. An ablation study found that excluding the in-vehicle features from the model reduces the model’s performance significantly, it could reduce the performance by up to 42%. In fact, the same experiment showed that having only the in-vehicle features is preferable to historical passenger flow features. Therefore, we conclude that using in-vehicle sensor data can be a feasible alternative to historical AFC data for predicting the passenger load.

The methodology has been extended by constructing a short-term forecasting model based on Seasonal ARIMA and GARCH that uses real-time signals of the passenger load to update its forecasts. The results show that while the forecasts lack accuracy initially, once the model is updated the forecasts improve up to almost a negligible error. When model predictions are used as update signals, the forecasting model is still able to improve and the results are competitive, despite the error contained in the signals.

Overall, we conclude that the proposed real-time model offers a suitable method for passenger load prediction and clearly demonstrates the effectiveness of using in-vehicle sensor data as input features. Moreover, we have presented a feasible method for using these features in a forecasting setting with a real-time model as an intermediary

A lot of models have been proposed to automatically complete code with promising evaluation results when tested in isolation on testing sets. This research aims to evaluate the performance of these models when used by developers when programming. Are these models still useful for actual programming and do developers even want this functionality? The model evaluated in this study is the InCoder model by Facebook, specifically the ability to complete code statements for the Python programming language. To evaluate this a plugin called Code4Me was made for PyCharm and VSC that will show code completion suggestions from the model when a keybind is pressed or a trigger point is encountered. If the user is shown a suggestion the plugin will send the actual line of code made by the developer after a delay, this can also be the suggestion itself if the user thought the suggestion was correct. When the user has used the model sufficiently they will also be asked to fill in a survey to gather opinions on the functionality that the model provides. The results show that there is a 21.95% ExactMatch, 52,73% edit similarity, and a BLEU-4 score of 36.05 for the statement completion functionality of InCoder. All users that filled in the survey preferred the automatic suggestions on trigger points but some indicated the keybind functionality was also useful. If the suggestion shown to the user was good they would use it instead of typing it themselves. The users indicated that the suggestions were a lot or a bit better than the default suggestions and using the plugin did save time programming. Overall all users were positive of the performance and thought the statement completion functionality provided useful suggestions.

State-of-the-art machine learning-based models provide automatic intelligent code completion based on large pre-trained language models. The theoretical accuracy of these models reaches 70%. However, the research on the practicality of these models is limited. Our paper will discuss the usefulness of UniXcoder, a machine learning-based cross-modal auto-completion model, in a normal environment through user evaluation. These models incorporate context around the requested completion and then return a prediction of code based on the context. To accomplish this, two plugins were made called 'Code4Me'. One for Visual Studio Code and PyCharm. These plugins work with a remote API that requires a segment of 3966 characters of the left and right context at the trigger point. The data collected consists of the inserted code completion, verification of the code completion, the IDE used, the trigger point, and the inference time. To evaluate the data the following metrics are used: BLUE~4, ROUGE-L, Exact Match, Edit Similarity, and the METEOR score. The results point out that developers accept once every 8 suggestions with an Exact Match of 62.5%, and the user evaluated, albeit with limited responses, are favourable towards the model and Code4Me. The accuracy of UniXcoder is lower in a real-world setting than when it is predicting source code. However, the usefulness of UniXcoder as an auto-completion model is apparent.

Code Completion is advancing constantly, with new research coming out all the time. One such advancement is CodeFill, which converts source files into token sequences for type prediction. To train the CodeFill model, a lot of source files are needed which take a long time to convert before training can begin. Converting the file the end-user is working on for completions is also essential for the total latency as longer files can affect the experience of using the model. In this study we aimed to improve the performance and success rate of this conversion. Our results indicate that we increased both the performance by 83 times or more depending on the input file length and the success rate by up to 45%.

Automatic code completions are a widely used feature when programming code efficiently. These completions can be made by various code language models, and these can be differentiated in three categories: single token completion, statement (line) completion and block completions. These completions, and in particular statement predictions are usually created using only the left context, missing key information and context on the other side. InCoder, a novel state of the art model is capable of using both contexts. In this study we aim to show the impact of using both contexts in statement completions. The results show that on average, an improvement of 9.9% exact match and similar results for Edit Similarity, BLEU-4, ROUGE-L F1, and METEOR when using both contexts instead of only the left context.