In recent years, strong progress has been made in creating learnable affine-equivariant models for downstream tasks such as classification. However, these models encounter increased data requirements to represent all possible transformations due to greater task complexity, while having been shown to generalize poorly to out-of-distribution data. In this work, we introduce a test-time approach for generalizing to out-of-distribution data. Namely, by utilizing a network trained to reconstruct any image that is part of a standardized training distribution, our model can infer an affine transform that moves new samples in-distribution by minimizing their reconstruction loss. With this, this approach closely matches the work of Spatial Transformer Networks, which instead learn to transform data, and inverted neural renderers for pose estimation. Through experiments, we show that this method contains a strong level of out-of-distribution translation and scale invariance, as well as a small level of rotation invariance. Namely, we show that it can handle significant transformations beyond those produced by commonly used benchmarks such as AffNIST. Using this strength we show that this method excels especially in low data regimes, outperforming existing competitors. ... Read more

Rendering shadows remains a fundamental problem in computer graphics, with existing approaches balancing performance against visual quality. Traditional shadow mapping techniques are efficient but suffer from aliasing artifacts due to its discretized nature. On the other hand, ray tracing can produce high quality shadow without discretization but is often too demanding for consumer hardware.

This thesis proposes a neural shadow representation that models a continuous mapping from light rays to occluder depth. Instead of relying on discretized buffers, we represent each object's shadow with a fully connected neural network. Given a ray origin and direction, the network predicts the depth at which the ray intersects geometry. This allows for depth tests similar to shadow mapping while supporting continuous input. The model is trained using ray-traced supervision with a dead-zone loss function that encourages the model to output depth corresponding to a position inside the occluder's geometry. It does not rely on discretized shadow maps and supports fully dynamic scenes, point lights, and directional lights.

Our results show that this method can achieve visual quality comparable to medium-resolution shadow maps while eliminating aliasing. The proposed method also has a unique ability to overfit to restricted light configurations. In scenes such as outdoor scenes where light movement is limited, the model's capacity is concentrated on relevant directions, allowing it to capture finer geometric details. While limitations remain regarding inference speed and high-frequency geometry, the proposed method demonstrates unique strengths that open up new trade-offs, particularly in scenarios with partially restrained lighting configurations and large scale but low-frequency geometry such as landscapes. ... Read more

Partial differential equations are widely used to model physical and geometrical behaviour. Real-time graphics and interactive simulation applications often require fast approximations of these equations. Traditional solvers often require solving large linear systems, which can form a bottleneck for real-time applications. Neural operators aim to learn the direct mapping between input and solution functions. This thesis investigates whether neural operator architectures can learn solution operators for partial differential equations on fixed triangular mesh surfaces. Several model families are compared, including a standard Multilayer Perceptron, DeepONet, a Graph Neural Network, a Multigrid Graph Neural Network, a Spectral Graph Neural Network, and a Hodge Spectral Graph Neural Network. These models are trained and evaluated on three surface-based PDE tasks: the Poisson equation with different scalar right-hand sides, geodesic distance approximation using the heat method with a varying number of source points, and linear elasticity using a Reissner–Mindlin thin shell with different vector-valued force fields. The predictions are evaluated using relative $L_2$ loss, isoline comparisons for scalar fields, error visualisations for displacement fields, and inference speed measurements. The results show that different PDEs favour different architectures. Spectral models perform especially well on the Poisson problem, particularly for smooth right-hand sides. Geodesic distance approximation is harder to learn, although the Hodge Spectral Graph Neural Network performs best overall for this task. The linear elasticity experiments show a different trend: for smoother force fields, the standard Multilayer Perceptron is highly competitive and often achieves the lowest loss, while localised force fields remain more difficult to approximate. This work also shows that neural operators can achieve significant speed-ups over traditional solvers, especially for more complex problems and larger meshes, suggesting a valid use case for these models in real-time applications. ... Read more

Self-supervised learning (SSL) lets computer vision models learn from unlabelled image datasets. Most DINO benchmarks pretrain on ImageNet — a million-image dataset that takes days of multi-GPU training per run, out of reach for the rapid iteration cycles smaller research groups rely on. This leaves practitioners with smaller datasets unsure whether DINO is worth running, or which of its design choices still hold at this scale.

We pretrain a small Vision Transformer (ViT-Tiny/8) using DINO on Tiny-ImageNet subsets from 1K to 100K images at 64x64 resolution, evaluated on downstream classification tasks. Downstream accuracy grows steadily with pretraining-set size and approaches the accuracy of a fully supervised baseline at the largest scale.

Our main contribution is a multi-crop ablation across data scale, training duration, and downstream task category. We find that multi-crop's benefit at sub-ImageNet scale is delayed rather than absent, and that the optimal multi-crop count depends on the downstream task category — no single setting wins across all tasks.

These findings show that the canonical DINO recipe does not transfer cleanly to sub-ImageNet scale. We recommend choosing the multi-crop count based on training budget and downstream task type, rather than copying the ImageNet default. ... Read more

Modern Automatic Speech Recognition (ASR) models show reduced performance on speech produced by children with Developmental Language Disorder (DLD) due to data scarcity, pronunciation variability and the different linguistic characteristics. This study investigates whether data augmentation, specifically converting Dutch dysarthric adult speech into Dutch child speech via monolingual Voice Conversion (VC), can improve ASR performance for Dutch children with DLD. This is motivated by overlapping acoustic characteristics between dysarthria and DLD. MeanVC, a zero-shot model was used to convert dysarthric adult speech into child speech. Subsequently Whisper-V3-Large was used to fine-tune on various configurations of synthetic VC speech and real Dutch DLD child speech. While models fine-tuned on VC speech outperformed the zero-shot baseline when real DLD data was unavailable, it was revealed that unconverted dysarthric speech outperformed the VC-generated speech, suggesting that the VC process might have difficulties keeping the necessarily speech characteristics. The study concludes that while dysarthric speech has relevant acoustic characteristics, this specific VC pipeline does not meaningfully improve ASR performance for DLD. ... Read more

This qualitative study explores how final-year Computer Science students perceive and manage conflicts during collaborative software projects. While these projects simulate professional teamwork, they frequently introduce interpersonal and organizational challenges. Using semi-structured interviews and Reflexive Thematic Analysis, this research investigates the subjective experiences and resolution strategies of students navigating team conflicts. The findings reveal three primary themes regarding conflict perception and resolution outcomes. First, teams generally manage minor task-oriented disagreements constructively through clear communication, resulting in incidental friction with no lasting impact. Second, prolonged behavioral conflicts, such as teammate absenteeism, can be resolved positively when managed constructively; these experiences often lead to personal growth, leadership, and communication skills. Third, escalating relationship conflicts, characterized by disruptive behavior and ineffective escalation processes, lead to significant negative consequences, including academic delays and emotional strain. The study concludes that while students demonstrate resilience and leverage routine teamwork obstacles to develop essential professional competencies, they struggle to navigate severe interpersonal conflicts without adequate external mediation. To optimize collaborative learning, software engineering curricula could integrate dedicated conflict management workshops, and universities must ensure accessible, transparent institutional support structures for complex interpersonal dynamics. ... Read more

Generative AI (GenAI) tools such as ChatGPT have become widely used by computer science students for code generation, debugging, and concept clarification. While these tools can improve efficiency in programming tasks, concerns remain about their potential impact on the independent problem-solving skills that computer science education seeks to cultivate. This study investigates how the use of GenAI during programming tasks shapes the way Computer Science students engage with independent problem-solving. Drawing on semi-structured interviews with seven Computer Science students at TU Delft, Through reflexive thematic analysis, the study identifies two themes. First, students framed their reliance on GenAI as a deliberate, self-governed choice rather than a loss of ability: they kept the interpretive work of understanding and planning, and the judgment of whether a solution was correct, for themselves, while delegating execution and The filling of knowledge gaps, and maintained that they could still work unaided if they chose to. Second, students reported a weakening willingness to persist through difficulty without help, both an immediate lowered threshold to begin effortful tasks and a gradual decline of motivation over time. Mapped onto Pólya's four stages of problem-solving, engagement is retained at the early and evaluative stages but surrendered in the middle, Persistence governs whether students return to a problem after a failed attempt. These findings suggest that GenAI reshapes independent problem-solving less by removing ability than by lowering the effort required to avoid it, and contribute to ongoing discussions about the responsible integration of AI tools in computer science education. ... Read more

This qualitative study explores how Bachelor's computer science students at TU Delft experience curriculum coherence across horizontal (concurrent) and vertical (sequential) dimensions. Based on semi-structured interviews analyzed using Reflexive Thematic Analysis, the results reveal how students view the trade-offs between academic depth and breadth, how well-timed repetition mitigates cognitive load, and how independence between parallel courses combats exhaustion. These insights highlight that curriculum designers should explicitly articulate connections between courses and carefully incorporate repetition to optimize student engagement. ... Read more

Interdisciplinarity is important for Computer Science education because it prepares students to solve real-world problems and collaborate with different disciplines. Despite its benefits, computer science students struggle with interdisciplinary learning. Without understanding the students’ motivations and struggles, it is difficult to design interdisciplinary courses that are beneficial and accessible. This research aims to understand how these students experience interdisciplinary courses, in terms of benefits and challenges. Seven computer science students were interviewed, and their transcripts were analysed using reflexive thematic analysis. Within interdisciplinary learning, the students experienced broadened perspectives and the communication of their ideas across disciplines. By understanding these experiences, faculty can better understand how to design courses centred around interdisciplinarity. ... Read more

Assessments are a crucial, high-stakes part of the learning process, and students consolidate much of their learning during the considerable effort they make preparing for them. Computer science curricula typically include both theory-heavy and programming-heavy assessments. This study investigates which learning strategies final-year computer science students adopt when preparing for theory-heavy assessments compared to programming-heavy assessments. Using semi-structured interviews with three final-year Computer Science and Engineering students at TU Delft, transcripts were analysed through an inductive open-coding pass followed by deductive coding using an established framework of cognitive and metacognitive learning strategies. Results show that rehearsal and elaboration strategies—reviewing lecture slides, note-taking, and creating cheat sheets—were predominantly associated with preparation for theory-heavy assessments, while metacognitive self-regulation, through consulting assignment specifications, was predominantly associated with preparation for programming-heavy assessments. Critical thinking was the one strategy reported across both assessment types, expressed through the use of practice exams. These findings suggest that practice exams function as a strategy supporting critical thinking regardless of assessment type. The relevance and diversity of practice exam material may therefore be an important, underexplored lever for supporting effective learning in computer science education. ... Read more

Edge AI applications, such as a wearable epileptic seizure monitor, demand efficient processing architectures. Spiking neural networks, computation-in-memory, and reservoir computing are neuromorphic concepts that come together in the Liquid State Machine (LSM), creating a promising path toward these efficient architectures. However, existing research on the LSM relies on idealized software simulations and simplified benchmark datasets, leaving the viability of practical real-time EEG classification unproven. In this work, a fully spiking LSM is proposed suitable for continuous real-time EEG classification. In a Brian2-based simulation framework, the LSM is first extensively analyzed with synthetic data, and then benchmarked using real electroencephalography (EEG) data. With a performance of 93.24% on the Bonn university dataset, and 77.13% on the CHB-MIT dataset, the proposed LSM outperforms all LSM literature. However, the experiments also revealed that the model's performance does not outperform a simple signal-intensity heuristic, and that its fading memory is mainly driven by short-term plasticity and higher-order dynamics in the synapses. Furthermore, the architecture exhibits high sensitivity to hardware non-idealities. Ultimately, a standard LSM is infeasible for EEG processing because its recurrent reservoir dynamics fail to outperform simple baselines. ... Read more

Self-supervised learning eliminates the need for image labels to learn meaningful visual representations, but it does not remove the need for large pretraining datasets. This work studies how Image-based Joint-Embedding Predictive Architecture (I-JEPA) behaves when pretraining data is deliberately limited. We train I-JEPA on stratified Tiny ImageNet subsets and evaluate the frozen representations with CIFAR-10 linear probing. The results show a steep improvement from the smallest subsets to the medium-data regime, followed by a plateau around the largest subsets under the standard final-checkpoint protocol. We also test two architectural modifications motivated by I-JEPA's design: reducing predictor capacity, to test whether an over-expressive predictor absorbs the pretext task instead of forcing useful encoder features, and adding shared photometric augmentation, to test whether extra input variation helps in low-data training. The shallow predictor improves transfer at 32k and 64k images but is neutral or harmful at the smallest and largest splits. The augmentation decreased downstream accuracy at 16k and was neutral at 32k. Additional controls---predictor depth sweeps, fixed-update budgets, and intermediate checkpoint analysis---suggest that the largest-split plateau is partly a training-dynamics issue rather than a pure data-efficiency ceiling. A cross-method comparison with Barlow Twins, MoCo, DINO, and MAE under the shared protocol contextualizes I-JEPA's data efficiency among SSL alternatives. ... Read more

Automatic Speech Recognition (ASR) systems perform poorly on speech of children with Developmental Language Disorder (DLD) due to a scarcity of training data and the high variability in speech of these children. This study investigates whether speech created through cross-lingual child-to-child voice conversion (VC) can improve ASR performance for Dutch children with DLD. To this end, multiple data augmentation pipelines were developed using the MeanVC zero-shot VC model, generating synthetic child speech with different source-target speaker pairing strategies (most similar, ran- dom, least similar) and varying augmentation sizes. The augmented datasets were used to fine-tune a Whisper-based ASR model. The results show that fine-tuning the ASR on a combination of synthetic speech and source speech improved ASR performance over an out-of-the-box baseline, but did not lead to better performance than fine-tuning it on real speech of children with DLD alone. Interestingly, speaker-dissimilar VC pairings yielded the best ASR performance. Subjective Mean Opinion Score (MOS) evaluations suggest that synthetic speech was perceived as less natural, less intelligible, and less child-like than real speech and that AMOS predictions only partially align with human judgments and ASR performance. These findings suggest that while cross-lingual voice-converted speech can have value as supplementary training data, improvements in speech quality and voice conversion methods are needed before it can serve as an effective replacement for real speech of a child with DLD in ASR training. ... Read more

Large language models (LLMs) can reproduce passages from their training data verbatim, raising privacy and copyright concerns. Prior work attributes memorization to factors such as model size, sequence entropy, context length, and repetition, but these findings lack a unified explanation. This thesis proposes a disambiguation complexity framework: memorization speed is governed not by the information content of a sequence, but by the difficulty of identifying it, specifically by the complexity of the minimal conditioning structure the model must extract from context to uniquely determine the correct continuation.

We demonstrate a counterintuitive regime in which random token sequences are memorized faster than structured natural language, contradicting standard explanations. We formalize a hierarchy of conditioning levels and introduce K-arity, a scalar complexity measure counting the number of prefix tokens jointly required to make a continuation deterministic. Through controlled experiments on synthetic datasets, we show that conditioning level and K-arity are predictive of memorization behavior. Attention analysis reveals that disambiguating cues are most clearly visible in early attention patterns. Natural language experiments show that, in text rich with redundant linguistic cues, isolated manipulations of conditioning complexity do not produce detectable differences, highlighting the gap between synthetic and naturalistic settings. This single principle connects input representation, entropy, identifying tokens, and context length within a common theoretical lens. ... Read more

Automated quality inspection of cluttered, multi-instance e-grocery stock containers is challenging because packaging clutter, product boundaries, container wear, and loose debris generate local anomaly responses that do not correspond to genuine quality issues. This thesis develops a modular reference-based inspection pipeline in which normality for each stock keeping unit (SKU) is modelled through automatically constructed and sampled reference sets. Query images are scored by patch-level nearest-neighbour matching of DINOv2 features against a reference memory bank, following the AnomalyDINO paradigm. Edge downweighting and crop-level debris filtering suppress boundary- and debris-driven false positives, while an optical character recognition (OCR)-based branch additionally addresses visually subtle wrong-SKU substitutions. Evaluated on 2,996 normal and 1,657 issue images across 20 SKU classes and five issue families, the full pipeline increases precision from 0.511 to 0.834 and reduces false positives from 1,414 to 201 relative to the core detector (F1-score: 0.705). This improvement comes at a recall cost, with recall decreasing from 0.893 to 0.610. The full pipeline achieves a 99th-percentile end-to-end latency of 1.4 s, within the 10 s operational constraint. The results demonstrate that robust inspection in this setting requires treating false-positive mitigation as a core system requirement rather than relying on raw anomaly sensitivity alone. ... Read more

This thesis presents an improved normal-guided pointcloud denoising pipeline that enhances the quality and efficiency of 3D pointcloud reconstruction. Building on the Constraint-based Point Set Denoising (CPSD) method by (Yadav et al., 2018), several modifications and extensions are proposed to improve the denoising process. The key contributions include a revised point classification approach, dedicated point update formulas for different point classes and a pipeline optimization and evaluation. Experiments were performed on synthetic and real-world scanned datasets, using Chamfer Distance (CD) and single-sided Chamfer Distance (sCD) as evaluation metrics. Results demonstrate that the proposed method achieves lower error scores than existing pipelines while requiring fewer iterations. Additionally, the modular nature of the pipeline enables future integration of neural networks or curvature-aware point update functions, opening pathways for further improvements in denoising pipelines. ... Read more

Large Language Models (LLMs) have demonstrated impressive capabilities on wide range of tasks including tasks that entail complex reasoning. They also demonstrate the ability to adapt to new tasks without further training but with the help of exemplars demonstrating how to solve the complex reasoning task. This is due to emergent capabilities such as In-Context Learning (ICL) where the model learns the skills required for a task through the demonstration samples provided. These methods can be categorized as static methods where exemplars are selected offline or dynamic (Instance-level) where exemplars are selected on a per test query basis. Dynamic, instance-level exemplar selection has been shown to be more accurate than static, task-level methods, but it is hard to use in practice because it requires a lot of computing power. In order to mitigate this issue we propose a novel perspective for selecting exemplars by casting it into a ranking problem and use LTR models trained on automatically generated BERTScore-based relevance labels to assign utility to the exemplars. However, randomly selecting and receiving llm feedback for exemplars may not yield the best data to train LTR models. Hence, principled exploration of the exemplar space is critical to learn a selection policy offline that can be easily employed for dynamic exemplar selection during inference. We tackle with these problems in this paper by proposing CASE Rank, a novel non-linear gap-index bandit framework that cuts down on inference-time overhead by learning an exemplar utility estimator offline without hurting performance. CASE Rank solves these problems by combining a gap-index based bandit framework and LTR using PiRank, a lightweight neural ranking model, as a non-linear surrogate loss function within the bandit framework. CASE Rank is a bandit based selection approach to judiciously sample LLM feedback and learn offline policy using a differentiable sorting algorithm. This approach allows for quick and tailored selection of exemplars for each instance during inference. Experiments conducted on datasets such as GSM8K, AQUA-RAT, and WMT19 indicate that CASE Rank enhances reasoning performance compared to previous methods, while also substantially lowering computational requirements. Our results highlight that principled, efficient exemplar selection can be achieved through a combination of exploration strategies and learning-to-rank models tailored to LLM response behavior. ... Read more

Coastal zones are dynamic and vulnerable regions, demanding accurate, scalable monitoring tools to inform environmental management and hazard mitigation. While satellite imagery and CNN-based classifiers have improved automated mapping, their reliance on unstructured pixel data limits contextual understanding. This study presents the first fine-tuning of a multi-modal large language model (MLLM), Qwen2.5, on 12-channel satellite input for multilabel coastal classification, demonstrating how architectural adaptation enables integration of spectral, topographic, and derived features beyond RGB. We compare this approach to a ResNet-50 baseline and state-of-the-art prompting methods using GPT-4o and LLaMA-3.2. Our experiments on the CoastBench dataset reveal that MLLMs benefit substantially from few-shot prompting with diverse, balanced sampling and that fine-tuning Qwen2.5 with full 12-channel input outperforms its RGB-only variant. An ablation study quantifies the importance of elevation and water-sensitive indices, while a human benchmark exposes a performance ceiling near F1 ≈ 0.70 due to label ambiguity. Our findings suggest that while MLLMs can rival traditional models and offer interpretability benefits, future gains depend on dataset quality, input diversity, and prompting strategy design. ... Read more

State-of-the-art models are susceptible to adversarial attacks. These attacks can cause catastrophic misclassification when robustness is required. With the increasing popularity of the retrieval augmentation paradigm in deep learning, we adopt it as a fully differential framework for adversarial robustness. We evaluate our method on three visual classification datasets, including ImageNet and attack our model with two white box attacks and a black box attack under various L₂ and L_∞ norms. The results indicate that a robust classifier emerges if the model fully relies on retrieved examples. We find that we can already obtain a PGD robust ImageNet classifier with 80.1% clean and 64.7% adversarial accuracy, using only one or two examples per class from the training data in the memory set. Contrary to other adversarial defense mechanisms, our method works directly on top of pre-trained models and remains robust when other defenses start to degrade for PGD attacks increasing in strength. ... Read more

High-resolution image analysis using deep Convolutional Neural Networks (CNNs) faces significant memory constraints due to the quadratic growth of intermediate feature maps with input resolution. This paper investigates whether Implicit Neural Representations (INRs), specifically SIRENs, can effectively represent CNN feature maps to reduce memory footprint during training. We address the unique challenge that CNN feature maps are not static signals but evolve continuously as network weights are updated through gradient-based optimization. Through three experiments on a modified All-CNN architecture trained on MNIST, we validate that: (1) SIRENs can fit static feature maps from frozen CNNs with high fidelity (PSNR > 30 dB) regardless of weight initialization; (2) SIRENs can track evolving feature maps during training, though with reduced reconstruction quality compared to static targets; and (3) SIREN-assisted feedforward—where SIRENs predict missing activations in receptive fields—enables classification accuracy (20.97%) above random guessing (10%) but substantially below standard training (95%). While results demonstrate the feasibility of using SIRENs to represent dynamic feature maps, significant challenges remain in maintaining reconstruction fidelity when SIRENs are integrated into the training loop. This proof-of-concept study provides empirical insights into bridging continuous implicit representations with discrete deep learning pipelines and highlights promising directions for future research in memory-efficient high-resolution image analysis. ... Read more