Accurate estimation of the noise power spectral density (PSD) is essential for effective single-channel speech enhancement in hearing aids, where stringent computational and latency constraints demand efficient noise estimation algorithms. This thesis investigates low-complexity noise PSD estimation techniques for system-on-chip (SoC)-based hearing aids, with a primary focus on minimum mean-square error (MMSE) estimation. After reviewing representative state-of-the-art signal-processing-based approaches, including minimum statistics, recursive averaging and MMSE-based estimators, an improved MMSE-based noise PSD estimation method is proposed to achieve a better trade-off between estimation accuracy, tracking capability, and computational complexity for resource-constrained hearing-aid platforms.

The proposed algorithm is evaluated under stationary and non-stationary noise conditions using synthetic, real, and speech-contaminated signals across multiple signal-to-noise ratios. Experimental results demonstrate improved noise tracking accuracy and reduced estimation variance while maintaining computational efficiency suitable for real-time embedded implementation. The proposed method provides an effective and practical solution for robust noise PSD estimation in next-generation hearing-aid systems.
... Read more

... Read more

Wave Function Collapse (WFC) is an algorithm for procedural content generation of tile-based images and game worlds out of adjacency constraints. It has gained traction over the past few years due to its ease of use and simplicity. However, slow runtime remains a major bottleneck of WFC, making it inconvenient to use for the generation of larger outputs. Having faster runtime would both allow for larger worlds to be generated in a reasonable amount of time, as well as make it practical for real-time use. Game developers would also be able to iterate more easily by having to wait less while designing. Using parallelisation would be an effective speedup method, but is nontrivial due to the sequential nature of the algorithm. The constraint propagation step of WFC, which is the most time-consuming step, is difficult to parallelise due to it being sequential in nature. Our solution to this is Binary Space Partitioned Wave Function Collapse (bspWFC), which generates the output image by following the structure of a BSP tree. Each time the tree splits the image in half, the border between the regions is generated, then the two regions can be generated separately in parallel. This process continues recursively up to a desired depth. As more and more branches of the tree are explored, more CPU threads can be utilised, until eventually they are all used at the same time. The individual tree nodes can also be rerun individually in the case of a conflict, improving the algorithm’s success rate. We implemented this algorithm in C++ and benchmarked it to evaluate its performance, measuring runtime, success rate and visual quality. A variety of tilesets were used to test the algorithm’s behaviour under different constraints, yielding significant improvements. For most tilesets, doubling the number of used threads resulted in about a ×1.7 runtime improvement. It also improved their success rate at large output sizes. Many tilesets that would normally have a successful generation less than 10% of the time were improved to over 90%. Due to the modified generation method, the output images could show noticeable artefacts, but this is only the case for a small number of tilesets. The biggest strength of our approach is that it allows parallelisation while remaining compatible with tilesets that would often propagate their constraints very far. ... Read more

Question answering (QA) systems can make information easier for children to access by providing direct answers to their questions. However, an answer that contains the correct information may still fail to meet a child's information need if its language is difficult to understand. Successful QA for children therefore depends on both the information provided and the linguistic form in which it is expressed. This thesis examines these two respective aspects through semantic adequacy and linguistic suitability.

To make linguistic suitability interpretable, the study first investigates how individual linguistic properties vary across texts intended for different levels of complexity. A broad set of linguistic measures is examined across two independently constructed collections of levelled texts, and the observed patterns are compared with expectations from the literature. This analysis identifies properties that vary meaningfully with level and provides an empirical basis for interpreting linguistic differences in QA responses. The retained properties are then used to evaluate responses to 1,400 grade-adapted science questions produced by four QA systems representing retrieval-based, extractive, and generative approaches. Semantic adequacy is assessed through complementary measures of correspondence with the expected content, while linguistic suitability is evaluated by comparing the linguistic profile of each response with the level-adapted ground truth for the same question.

The first analysis shows that several linguistic properties vary systematically with increasing text complexity, while others show weaker or less consistent relationships. Most properties for which the literature provides a directional expectation follow that expected direction, although some do not. In the QA evaluation, the generative systems show the strongest overall performance, although their semantic profiles differ across measures. The retrieval-based system performs less strongly overall, while the extractive system shows the weakest profile. The feature-level linguistic analysis further shows that different properties can account for a system's distance from the ground truth at different grade levels, even when its overall distance changes little.

The results show that evaluating QA responses for children benefits from considering both semantic adequacy and linguistic suitability. By decomposing linguistic profile differences into interpretable properties, the proposed approach identifies not only how closely system responses align with level-adapted answers, but also the linguistic characteristics through which that alignment or divergence arises. ... Read more

As machine learning becomes a standard part of computing and engineering curricula, teaching its core concepts effectively has become an important educational challenge. Regularization is one such concept: L1 and L2 penalties are widely used to prevent overfitting, with L1 producing sparse solutions and L2 shrinking weights more smoothly. It is commonly taught through two representational formats: an algebraic one, presenting the loss function and its penalty terms, and a geometric one, depicting constraint regions and their intersection with the loss contours. However, instructors choose between them without clear evidence on which better supports learning. This study asks whether teaching regularization algebraically or geometrically leads to different student performance, and in which kinds of understanding.

Two interactive notebooks, matched on learning objectives, length, and reading difficulty, were developed to teach the concept in each format and are released openly; they were compared in a between-subjects experiment with students who had completed an introductory machine learning course. Learning was measured with a post-test spanning conceptual understanding, problem-solving, and knowledge transfer, alongside a thematic analysis of students’ written explanations.

Both formats supported practical reasoning about regularization, but not identically: the algebraic group performed better overall, with its clearest advantage in conceptual understanding, no reliable difference on problem-solving, and an inconclusive result on transfer. The two groups largely shared the same core understanding but expressed it through different vocabularies: a penalty on the loss function versus a shrinking constraint region in weight space. Representational choice therefore appears to shape how students explain regularization more than whether they grasp it, suggesting that combining the two formats may best support learning.
... Read more

Machine learning is increasingly important in computer science education, but introductory concepts can be difficult because they combine mathematical notation, algorithmic reasoning, and conceptual understanding. Gradient descent is one such concept: students may reproduce the update rule while still struggling to explain the role of the loss function, gradient, learning rate, and repeated parameter updates.

This paper investigates whether multiple representations can support beginner understanding compared with a classic textbook-style explanation.

A small-scale exploratory experiment was conducted with students who had little or no prior machine learning experience. Participants completed a prerequisite pre-test, studied gradient descent using either a text-and-formula-based explanation or a multiple-representations explanation, completed a post-test, and answered an experience survey.

The multiple-representations condition showed higher post-test performance, especially on computation and application tasks, as well as higher confidence, clarity, usefulness, and engagement. Perceived cognitive load remained similar across conditions.

These findings suggest that aligned multiple representations can help beginners connect formal notation with concrete calculations and intuitive understanding, although the results should be interpreted cautiously because of the small sample size. ... Read more

Machine Learning (ML) is a rapidly growing field within Artificial Intelligence and one of the most prominent areas of technological study, which is particularly challenging for new learners as it requires a strong grasp of abstract algorithmic structures along with rigorous mathematical reasoning. Despite this, traditional instructional approaches often fail to support deep conceptual understanding, particularly for foundational models such as Decision Trees. This study examines whether integrating multiple instructional representations (including textual explanations, visualizations, analogies, videos, and interactive simulations) enhances student learning outcomes compared to traditional text-only materials.

To examine the lack of empirical evidence on multi-representational teaching for Decision Trees in ML education, a mixed-methods pilot experiment was employed with 10 participants was employed, comparing a multi-representation tutorial group to a text-only group. After a pre-test on mathematical and logical reasoning, participants completed a structured learning phase and a post-test measuring conceptual understanding, problem-solving, and transfer. Semi-structured interviews were also conducted to capture learner experiences.

Quantitative analysis included independent group comparisons using descriptive statistics, and inferential tests. Qualitative data were analyzed with inductive and deductive thematic analysis.

This study provides preliminary evidence that multi-representational instruction may improve problem-solving performance in Decision Tree learning. However, due to the small sample size and lack of statistical significance on two of three outcomes, these findings should be interpreted cautiously and require replication with larger samples.

The study contributes a structured evaluation framework for multi-representational ML education and provides evidence supporting the potential benefits of interactive instructional design in teaching Decision Trees. ... Read more

Machine Learning has become a prominent component of STEM curricula, yet there remains a lack of standardized pedagogical frameworks for teaching it. Principal Component Analysis represents a typical educational challenge, requiring students to simultaneously coordinate algebraic manipulation, geometric intuition, and algorithmic thinking. This study evaluates how different combinations of instructional representations affect undergraduate students’ conceptual understanding, problem-solving performance, and knowledge transfer ability. An experiment was conducted with 27 first-year Computer Science students at TU Delft. Participants were assigned to one of two experimental conditions: a traditional static representations group or a multimedia-enhanced representations group. Quantitative analysis revealed that the static group significantly outperformed the interactive group in total post-test scores and knowledge transfer. No statistically significant differences were observed in conceptual understanding or problem-solving performance. Qualitative thematic analysis indicated a disconnect between perceived and actual learning: while students preferred the interactive widgets for building geometric intuition, these features may have introduced extraneous cognitive load, provided a false sense of understanding, or students simply ran out of time because interactive exploration takes longer. ... Read more

Recent text-to-image (T2I) diffusion models can generate highly realistic images, but they often struggle to correctly arrange multiple objects according to specified spatial relationships. This limitation reduces their usefulness as controllable design tools. The problem is particularly challenging for modern multi-modal diffusion transformers (MM-DiTs), such as Stable Diffusion 3.5 and FLUX, whose architecture prevents the direct application of earlier layout-control techniques. Existing solutions either require costly model retraining or use training-free methods that provide limited and often unreliable control. This thesis introduces FOCAL, a training-free layout controller that formulates spatial guidance as a stochastic optimal control problem during diffusion sampling. By applying a closed-form correction derived from the model’s attention maps, FOCAL simultaneously enforces object placement and attention separation without modifying model weights. The method improves compositional accuracy across multiple backbones and achieves performance competitive with much larger state-of-the-art systems. ... Read more

Misinformation on the web has been identified as one of the major dangers of the current digital society.
While it can affect many domains, the consequences are especially severe in certain contexts. One such context is health, where exposure to misinformation may lead to harmful health decisions. Because web search is a widely used tool to access health information, it is also an important context in which such misinformation is encountered.

All search processes begin with a query, and the way a query is phrased directly impacts the search results.
Query reformulation techniques have been used to modify queries, with the goal of improving search outcomes.
A large body of research has focused on developing different reformulation techniques that improve topical relevance. However, how such techniques affect the presence of misinformation in health search results remains unexplored. In this work, we address this research gap by conducting an empirical exploration of how different query reformulation techniques affect misinformation retrieval in health search. We consider a range of reformulation techniques based on different underlying methods and compare them on the TREC 2021 Health Misinformation dataset by analyzing their effects on misinformation, topical relevance, and correct information across ranked search results.

Our results indicate that "one size fits all'' is not the way to go, as most reformulation techniques do not systematically alter misinformation presence. However, we identify three techniques: two simple knowledge model-based approaches and one more recent LLM-based approach, that show potential to reduce misinformation relative to topically relevant information. While query reformulation cannot eradicate misinformation entirely, it can have an impact. Our research demonstrates both the potential of query reformulation as one component of a broader strategy for combating misinformation, and the importance of considering misinformation in the future development of reformulation techniques. ... Read more

Millimeter-wave (mmWave) radar is a promising active sensing technology for Human Pose Estimation (HPE). However, its reliability is hindered by poor generalization in scenarios unseen during the model's training phase. This thesis presents a comprehensive empirical study to analyze and quantify the mmWave dimensions causing the poor generalization for HPE models. To enable this research, we introduce mmDiverse, a new large-scale dataset containing varied human movements, users, environments, and distances from the radar. Using this dataset, we evaluated two foundational models, Baseline MARS and a temporally aware version, Temporal MARS, through a series of experiments designed to isolate each dimension. The results reveal that human diversity is the most critical challenge, with model accuracy dropping by over 100% when encountering an unseen individual. Unseen movements pose the next significant challenge, where the models revert to the learned movements exposed during model training rather than generalizing to new kinematic movement patterns. Additionally, our study shows that the model's capacity to learn new kinematic patterns is enhanced by integrating model-centric techniques such as temporal modeling. This study also reveals that training in cluttered, noisy environments, combined with a target classification and tracking data pre-processing pipeline, is crucial for improving model robustness. Based on these findings, this thesis provides a set of evidence-based guidelines for developing more resilient mmWave-based HPE systems. This study concludes that building reliable applications requires prioritizing the collection of data with extensive user and movement diversity, captured across noisy and cluttered real-world environments. Furthermore, this diverse dataset should be used to train a temporal-aware model architecture incorporating a data pre-processing pipeline to mitigate generalization challenges. ... Read more

Test-time adaptation methods assume privileged access to model internals: parameters for fine-tuning, statistics for recalibration, or architectural components for modification. This assumption fails when models are deployed as certified systems, encrypted services, or under regulatory constraints that prohibit parameter changes. We present ITEM (Input Transformation via Entropy Minimization), the first preparation-agnostic sample adaptation method for test-time adaptation. ITEM learns input transformations that minimize prediction entropy using only gradient signals through frozen models, exploiting the principle that well-calibrated models produce low-entropy outputs on familiar data. Unlike existing sample adaptation methods that require specialized training procedures or parameter updates, ITEM works with any pre-trained model without modification or preparation requirements. Using scalar transformations as proof of concept, we demonstrate adaptation under extreme data scarcity: models trained on 10 samples per class and adapted with single calibration samples. ITEM significantly reduces performance degradation from corruption while existing methods fail or show negligible improvement. Our results establish that effective test-time adaptation is possible without model modification, architectural knowledge, or training preparation, opening new possibilities for adapting deployed models under real-world constraints. ... Read more

We explore the use of biologically inspired auditory kernels—learned from sparse coding on (clean) read speech—to analyze and reconstruct signals degraded with additive noise. Auditory kernels mimic spectrotemporal filters in the human auditory system, offering insight into how structured acoustic signals can be internally represented and selectively preserved. Our study applies an auditory kernel-based matching pursuit reconstruction framework to clean, degraded, and standalone noise audio, investigating kernel activation patterns across input types. The findings reveal kernel selectivity; structured signals like speech activate a common subset of kernels, while unstructured noise elicits distinct, less overlapping activations, allowing for more effective separation and implicit denoising. This selectivity results in implicit denoising, preserving intelligibility and perceptual quality even under degradation. By quantifying this behavior across noise types and SNR levels, we show that auditory kernels not only support robust signal reconstruction but also offer a biologically grounded, explainable mechanism for speech enhancement. These insights advance the use of sparse auditory models in both neuroscience and signal processing, motivating future work on adaptive or context-aware dictionaries. ... Read more

A persistent gender gap exists in computer science occupations which may be partly driven by the subtle use of gendered language in job advertisements and the biased behavior of job recommendation algorithms. This thesis investigates how gender bias in job ads has evolved over time and how it varies across countries, industries, job roles, and work models. It also explores whether recommendation systems expose applicants differently to masculine- or feminine-coded jobs based on gender. A dataset of nearly 470,000 LinkedIn job advertisements related to computer science from 2014 to 2024 was scraped, filtered, and labelled using a gender bias score based on a curated repository of gender-coded words. Synthetic CVs were generated to simulate male and female applicants and used as users for whom job recommendations were generated using five content-based recommendation models (TF-IDF and Word2Vec variants). Results showed that job advertisements are predominantly masculine-coded, though a decline in masculine bias has occurred since 2018. Variation in bias across country, industry, and role is statistically insignificant with low practical effect sizes. The TF-IDF model exhibited the highest disparity in job exposure, while Word2Vec-SkipGram showed more balanced recommendations. A weak correlation was found between applicant gender and the genderedness of recommended jobs. ... Read more

Autism Spectrum Disorder (ASD) is a neuro-developmental disorder reported to affect around 58 million people across the world. Due to their struggles with social communication in real life, they prefer to use online mediums to form social connections and are better able to maintain meaningful relationships through their communications over these online mediums. However, empirical studies have revealed that autistic users are not as efficient in using these online mediums when they have to search for information. While studies in the past have resulted in a growing list of guidelines to improve the web search experience of autistic users, whether existing online ISTs adhere to these guidelines has not been studied extensively. Thus while we know what the autistic users need to efficiently look for information on the web, we do not know if and how the ISTs are catering to their accessibility needs.

To advance knowledge in the accessibility of popular online ISTs when catering to the information needs of autistic users, we conducted an empirical exploration focusing on the accessibility of the textual
information in the IST responses. We specifically examined the responses on three aspects of text accessibility for autistic users - (1) text structure, (2) text readability, and (3) text concreteness using
“accessibility indicators” i.e. quantifiable features created by us based on widely accepted web content accessibility guidelines for autistic users. Due to the lack of standard query datasets for autistic users, we had to generate synthetic queries representative of autistic users’ information needs to collect responses from four popular ISTs Google, Bing, Gemini, and ChatGPT. We examined the differences in the way ISTs respond to queries characteristic to autistic users in contrast to queries typically asked by the general public, as well as the differences in IST responses across different kinds of queries asked by autistic users. We also investigated if an autistic user were to reformulate their query and explicitly state that they are autistic in their query would nudge the IST to respond differently to their query.

Juxtaposing the IST responses generated for control and ASD group queries revealed that the ISTs not only responded differently to the queries asked by the two user groups, but the responses generated for the control group were surprisingly more accessible for autistic users than the IST responses generated for their queries. For queries typically asked by autistic users, none of the ISTs produced responses
suitable for autistic users. However, each IST had its strengths in the context of the accessibility which can employed in the other ISTs to improve the general accessibility of the responses generated by all ISTs. We also found that our query and prompt reformulation strategies affected the accessibility of each IST differently. In general, the reformulations did nudge the ISTs to generate responses that are more suitable for autistic users. This presents an optimistic solution to the issue of ISTs not catering to the accessibility needs of autistic users when they search for information on the web. Our results highlight areas where popular ISTs can be improved to make their responses more accessible to autistic users. Our systematic empirical investigation pipeline can also be extended to investigate IST responses on other factors that influence the accessibility of web content for autistic users. ... Read more

Indoor localization technology has become increasingly crucial as the demand for precise and reliable positioning systems grows across various applications. Traditional methods, such as vision-based techniques, radio signal-based technologies (including UWB, WiFi, RFID, and Bluetooth), and visible light-based technologies, offer unique advantages and limitations. Among these, visible light positioning (VLP) stands out for its potential to provide high accuracy by leveraging the characteristics of light signals.

This thesis explores the integration of VLP with a balloon-enabled drone—a novel UAV setup featuring a buoyant balloon that extends flight duration. A balloon-enabled drone introduces both opportunities and challenges for VLP methods due to its size. Its large surface area can block light paths, which may impact signal reception and positioning accuracy. On the other hand, it also allows for the use of multiple receivers across the surface, potentially improving positioning reliability.

Traditional VLP systems typically utilize multiple transmitters and a single receiver; however, our approach takes advantage of the large surface area of a balloon-enabled drone by using only a single transmitter with multiple receivers strategically positioned on the balloon. This setup leverages the balloon’s curved surface to capture a diverse range of light intensities and angles, thereby improving positioning accuracy. We developed a 2D+H RSS-based VLP model specifically designed for balloon-enabled drones. This model takes into account factors like light transmission and optical channel loss. Our VLP system includes multiple receivers placed on the balloon’s surface and a single transmitter. We analyzed the optimal number and placement of these receivers to enhance positioning accuracy.

The system’s performance was tested through both static and dynamic experiments. In static tests, our configuration achieved an average positioning error of 4 cm. During dynamic tests, which involved movement and tilt, the mean error increased to 10-12 cm, largely due to difficulties in estimating height and managing tilt angles. Overall, our system shows an improvement over existing positioning methods like Crazyflie, while also maintaining low energy consumption and computational complexity. This work highlights the potential of our VLP model to improve the positioning accuracy of balloon-enabled drones for various applications. ... Read more

This thesis involves the development and testing of an educational tool named ”Playlist,” aimed at improving the learning experience of Computer Science & Engineering bachelor students at the TU Delft. The Playlist tool was developed as a feature for the existing tool called Skill Circuits, supporting students in creating personalized sequences of learning activities for their study sessions. This research contributes to the complexities of modern digital learning environments and the challenges students face in navigating vast amounts of educational content.

This report holds the full description of the integrated thesis. The first part including the chapters of ”Fostering a personal sequence of learning activities in Skill Circuits”.
The study uses a design-based research approach, which involves iterative development and research cycles and including students in the evaluation phases. By drawing on educational theories and addressing practical needs identified at TU Delft, the Playlist tool was designed to seamlessly integrate into the existing learning environment. The tool’s functionalities were prioritized based on user feedback provided by Computer Science students and gathered through surveys and focus group sessions.

Key findings indicate that playlists can improve student engagement, motivation, and overall learning experience by supporting students in finding learning activities aligned with their learning preferences and goals. The research highlights the importance of providing students with tools that improve navigating the learning environment and support them in self-directed learning. The Playlist feature not only aids students in organizing their learning activities but also promotes a holistic approach to education by incorporating supporting activities such as reflection and taking breaks.

The integrated thesis hopes to demonstrate the potential and importance of designing tools to personalize learning and support students in achieving their learning goals. It also explains why personalization tools in higher education should be implemented in an iterative manner. ... Read more

A Room Impulse Response (RIR) is a mathematical model for sound propagation in a room. Estimating RIR parameters such as the reverberation time (T60) allows Automatic Speech Recognition (ASR) systems to adapt to reverberation in input signals by changing their behavior based on these estimates. Currently, machine learning techniques provide the most accurate T60 estimations. We propose a novel methodology by using intrusive Speech Intelligibility Measures (SIMs) beyond their traditional application. In this study we utilize SIIB, SIIB^Gauss, STOI and ESTOI as SIMs. For each SIM we find a best fit curve with respect to the reverberation time (T60) using a statistical approach. The statistical analysis is applied on simulated RIRs obtained by using the Image Source Method. The estimator for SIIB^Gauss achieves the lowest Mean Squared Error of 0.353 on simulated data. Although this does not outperform state-of-the-art models, we offer recommendations for possible improvements. Preliminary experiments suggest that enhancing noise robustness is crucial and that the estimators could be generalized to real-world scenarios. However, further research is necessary to confirm this. ... Read more