CS Theses & Dissertations 2025

For 2025 graduation dates (in alphabetical order by last name):

Multimodal understanding of long documents : from topic modeling to question answering
Abaskohi, Amirhossein
DOI : 10.14288/1.0449900
URI : http://hdl.handle.net/2429/92096 
Degree : Master of Science – MSc       
Graduation Date : 2025-11
Supervisor : Dr. Giuseppe Carenini

Budget-robust active learning
Bae, Wonho
DOI : 10.14288/1.0449756
URI : http://hdl.handle.net/2429/91949 
Degree : Doctor of Philosophy – PhD 
Graduation Date : 2025-11
Supervisor : Dr. Danica Sutherland

Deep learning has made significant strides in recent years, largely due to the availability of vast amounts of labeled data. However, expensive and time-consuming manual annotation limits the widespread adoption of Artificial Intelligence (AI), particularly for smaller organizations and individuals. This highlights the need for data-efficient AI frameworks that reduce dependence on large-labeled datasets, making AI more accessible. Active learning, where a model strategically selects the most informative data points for annotation, offers a promising solution to this challenge. It improves model performance with fewer labeled examples, making it especially valuable in domains where labeling is costly. Recent research has revealed that the effectiveness of active learning methods varies significantly across different budget regimes, where the budget is defined by the size of a labeled set. In particular, uncertainty-based methods, which perform well in high-budget settings, often underperform compared to representation-based methods or even random sampling in low-budget regimes. In this thesis, we investigate how to improve active learning under both high- and low-budget regimes. We begin with the high-budget setting, where we introduce a novel uncertainty-based method that leverages neural tangent kernels (NTKs) to make computation of look-ahead acquisition strategies feasible. This approach allows the model to account for the changes of ``future" predictions, resulting in strong performance across various datasets, particularly in the high-budget regimes. In the low-budget regimes, we propose MaxHerding, a representation-based method that generalizes the recently introduced ProbCover and establishes connections to other low-budget active learning techniques. To further explore active learning under limited annotation budgets, we consider its application to meta-learning (or few-shot learning) and develop a simple yet effective acquisition strategy based on Gaussian Mixture Models (GMMs), motivated by a max-margin classifier. Given the difficulty of determining the appropriate budget regime in advance, we finally propose Uncertainty Herding (UHerding), a budget-robust active learning method that adaptively interpolates between uncertainty- and representation-based strategies. Our empirical results show that UHerding consistently outperforms existing methods across a wide range of budget regimes, offering a promising direction toward hyperparameter-free and more robust active learning in real-world applications.

ML-empowered combinatorial search
Cameron, Christopher
DOI : 10.14288/1.0448579
URI : http://hdl.handle.net/2429/90814 
Degree : Doctor of Philosophy – PhD 
Graduation Date : 2025-05 
Supervisor : Dr. Kevin Leyton-Brown

Combinatorial search problems underlie a wide range of computational tasks, from resource allocation in 5G networks and spectrum auctions to verification of hardware and software systems. Despite the best-known techniques having worst-case exponential complexity, many real-world instances exhibit rich structural patterns that specialized solvers can exploit. No single solver excels across all problem distributions, and tuning solvers to specific distributions is crucial for best performance. When done by hand, this application-specific algorithm-design process is very tedious and limited. This thesis explores how machine learning (ML) can empower combinatorial search by integrating data-driven methods with symbolic solvers, maintaining correctness guarantees while adapting to specific applications. We begin by exploring evaluation and applications of two classic ML approaches for algorithm design: algorithm selection and configuration. We later relax two limitations of these classic approaches: (1) reliance on hand-crafted features to represent problem instances; and (2) a black-box view of one or more solvers. We relax the first constraint by learning an end-to-end neural representation of problem instances that captures the right invariances for reasoning about Boolean logic. We relax the second constraint by integrating ML into the solver’s internal decisions, proposing a new reinforcement learning approach, Monte Carlo Forest Search (MCFS), to learn improved branching policies in tree search. With these new advances, learning models for each new application typically demands extensive training data. We show it is possible to train foundation models that dramatically reduce the overhead of learning new tasks and distributions. We trained a single model on a large dataset across a broad range of combinatorial tasks that generalized both on held-out tasks and distributions. We found that models fine-tuned from the foundation models were substantially more sample efficient than models trained from scratch. Finally, we consider the predict-then-optimize paradigm, showing that coupling predictive models with the solver objective can prevent large degradations in solution quality. We formalize this in the setting where forecasts are inherently stochastic and multiple correlated forecasts drive downstream decisions.

Field of values analysis that includes the origin for preconditioned nonsymmetric saddle-point systems
Chen, Hao
DOI : 10.14288/1.0448490
URI : http://hdl.handle.net/2429/90701 
Degree : Master of Science – MSc 
Graduation Date : 2025-05 
Supervisor : Dr. Chen Greif

Thinking still matters: prompting reflection in software engineering education
Chin, Kyle
DOI : 10.14288/1.0449887
URI : http://hdl.handle.net/2429/92085 
Degree : Master of Science – MSc    
Graduation Date : 2025-11
Supervisor : Dr. Reid Holmes

Comparing the persuasiveness of humans and large language models in persona-based dialogue
Chockkalingam, Shruthi
DOI : 10.14288/1.0449140
URI : http://hdl.handle.net/2429/91377 
Degree : Master of Science – MSc  
Graduation Date : 2025-11
Supervisor : Dr. Vered Shwartz, Dr. Raymond Ng

Vision and language : representation learning, commonsense reasoning, and consistency
Chou, Shih-Han
DOI : 10.14288/1.0449876
URI : http://hdl.handle.net/2429/92071 
Degree : Doctor of Philosophy – PhD 
Graduation Date : 2025-11
Supervisor : Dr. Leonid Sigal, Dr. Jim Little

Vision-Language Models (VLMs), a prominent sub-class of multimodal architectures, are important because they allow automated systems to process and understand both visual and linguistic data, enabling more intuitive and accessible interactions. They bridge a gap between language and vision, leading to advancements in areas like visual search, virtual assistants and many others. The recent advent of Large Language Models (LLMs) has significantly accelerated progress in this domain, prompting new research directions and challenges. This thesis investigates vision-language modeling from several complementary perspectives, with the overarching aim of improving alignment, reasoning, and consistency in multimodal systems. First, we address the problem of fine-grained alignment between vision and language. We propose a semi-supervised grounding mechanism that leverages a pre-trained phrase grounding model to generate region-phrase pseudo-labels. This approach facilitates more granular alignment and enhances feature learning without requiring additional human annotations, particularly in large-scale settings. Second, we explore the integration of commonsense knowledge in multi-sentence video captioning. We introduce a Transformer-based model that incorporates both implicit (visuo-lingual and purely linguistic) and explicit (knowledge-base) commonsense information, which demonstrates that incorporating prior knowledge leads to more coherent and contextually appropriate captions. Third, we present a systematic analysis of the semantic consistency of Vision-Language Models (VLMs). We construct a benchmark comprising three tasks, Question Rephrasing, Image Restyling, and Context Reasoning, and evaluate model responses using correctness and consistency metrics. Finally, we propose a test-time adaptation framework to improve semantic consistency without requiring supervised re-training. This model-agnostic method optimizes two complementary objectives: a cross-entropy agreement loss to align predictions across semantically equivalent inputs, and a pseudo-label consistency loss that steers predictions toward a stable consensus. The method operates post hoc and uses only the test input, making it broadly applicable. Collectively, the contributions of this thesis advance both the methodological foundations and empirical understanding of multimodal learning, particularly in enhancing alignment, reasoning, and reliability in VLMs.

Towards efficient machine learning management systems 
Ding, Dujian 
DOI : 10.14288/1.0449274
URI : http://hdl.handle.net/2429/91495 
Degree : Doctor of Philosophy – PhD 
Graduation Date : 2025-11
Supervisor : Dr. Laks Lakshmanan 

Machine learning (ML), especially deep learning (DL), has become a leading force in both academic research and industrial applications. In tandem with the impressive capability of ML models, the model sizes have drastically exploded over recent years. Gigantic model sizes not only lead to computational challenges but also ethical concerns such as green AI and democratizing AI. Significant research efforts have been invested into making ML more efficient. Model-level ML efficiency studies the fundamental trade-off between model efficiency and effectiveness, while system-level ML efficiency deals with the overall efficiency of answering ML inference queries invoking multiple models. In this dissertation, we aim to answer the central question: How to make machine learning service more efficient without compromising the overall performance? Our work is aligned with both model-level and system-level ML efficiency. On the model level, we propose effective approaches to extract efficient subnetworks from gigantic ML models with user-specified sparsity targets while maintaining high performance (Chapter 2). On the system level, we identify two important sub-tasks: bulk query processing and streaming query processing, where query objects are either provided all at once as a bulk, or they stream in. As to bulk query processing, we study the Fixed-Radius Near Neighbour query and develop approximate algorithms to efficiently deliver high quality answers with statistical guarantees using one cheap proxy model and one expensive oracle model (Chapter 3). Additionally, we investigate the more general setup where we have a set of ML models at different cost and accuracy trade-offs and conceive principled algorithms to select the optimal model assignments for ML classification queries (Chapter 4). As for streaming query processing, we consider powerful conversational ML services such as ChatGPT which are supported by the development of powerful large language models (LLM). We develop novel adaptive query routing solutions to significantly reduce the overall incurred costs by distributing query traffic from the expensive cloud models to the small on-device models without compromising the response quality (Chapter 5). Furthermore, we extend the routing framework to a spectrum of LLMs and leverage best-of-n sampling techniques to further enhance efficiency (Chapter 6).

Precision-cascading in restarted GMRES
Dos Remedios, Brandon
DOI : 10.14288/1.0449657
URI : http://hdl.handle.net/2429/91891 
Degree : Master of Science – MSc 
Graduation Date : 2025-11
Supervisor : Dr. Chen Greif

Learning in networks with communication delays
Fayyazi, Ryan
DOI : 10.14288/1.0447408
URI : http://hdl.handle.net/2429/89783 
Degree : Master of Science – MSc 
Graduation Date : 2025-05
Supervisor : Dr. Frank Wood

Data-efficient learning on structured output data
Goyal, Raghav
DOI : 10.14288/1.0447394
URI : http://hdl.handle.net/2429/89767 
Degree : Doctor of Philosophy – PhD 
Graduation Date : 2025-05
Supervisor : Dr. Leonid Sigal

Deep learning relies on huge amounts of labeled data which is time-consuming and expensive to gather and annotate. The issue becomes even more pressing when dealing with structured data, where the cost of annotation increases with complexity of annotations from scribbles, bounding boxes, segmentation masks to scene-graphs. Efficient learning approaches attempts to alleviate this issue by utilizing lower quantity (em e.g., few-shot learning) and quality (em e.g., weakly-supervised learning) of annotations for a target task. In this thesis, we explore and develop efficient learning approaches to tackle structured output tasks across images and videos. Specifically, we first propose a unifying approach for any-shot (zero-shot and few-shot) object detection and segmentation using a semi-supervised transfer learning methodology which learns to semantically transform weak detectors/segmentors to strong ones. We then follow up on more granular annotations - scene graphs - and propose a simple weakly-supervised approach for human-centric scene-graph detection, where despite assuming weaker supervision for objects and relations, we perform competitively to state-of-the-art approaches. We then turn our focus to videos, which compared to images, are more label intensive due to an additional temporal dimension. We explore model design choices for videos in the context of efficient learning paradigms. In particular, we first look at dynamic spatio-temporal annotations in videos, where we propose a single, unified model for tackling multi-modal, query-based video understanding in long-form videos, and show that multi-task training leads to improved performance and ability to generalize to unseen tasks. And second, in the context of Long-Video Object Segmentation, we propose a transformation-aware loss which places greater emphasis on parts of videos where the tracked object is undergoing deformation, and show improved performance over prior works, along with a time-coded memory beyond vanilla additive positional encoding, which helps propagate context across long videos.

Factoring over probability simplices : from theory to applications
Graham, Naomi
DOI : 10.14288/1.0449899
URI : http://hdl.handle.net/2429/92099 
Degree : Doctor of Philosophy – PhD 
Graduation Date : 2025-11
Supervisor : Dr. Michael Friedlander

Non-negative matrix factorization (NMF) is a fundamental technique in modern data analysis. Its distinguishing characteristic—requiring non-negative factors—ensures that it models only additive components, producing interpretable parts-based representations. When augmented with sum-to-one constraints, yielding stochastic matrix factorization (SMF), this interpretability becomes mathematically precise through an exact correspondence with latent variable models. This thesis develops a comprehensive framework extending these ideas to continuous multivariate probability densities via non-negative simplex-constrained tensor factorizations. We demonstrate the framework’s versatility through two novel applications: sediment provenance tracing in geology and spatial transcriptomics denoising in biology. Algorithmically, we develop a novel rescaled block coordinate descent method that efficiently handles simplex constraints while maintaining iterate bound- edness—a key challenge in constrained NMF. By exploiting the semialgebraic structure of our objectives through the Kurdyka-Łojasiewicz (KL) property, we establish subsequential and full-sequence convergence guarantees. This analysis framework proves remarkably versatile: we successfully apply it to both our probabilistic tensor factorization models and to an adaptive graph-regularized denoising method for spatial transcriptomics, where the regularization evolves with the solution. Theoretically, we investigate why NMF methods work well despite non-convexity. Drawing inspiration from the benign geometry of low-rank matri- ces, we analyze the structure of SMF feasible sets through novel reparameter- izations and boundary characterizations. Our analysis reveals that spurious stationary points primarily occur at boundaries, suggesting new algorithmic strategies for avoiding local minima. These geometric insights, combined with our convergence guarantees and practical applications, provide a unified view of constrained factorization methods that bridges theory and practice.

TABI : tight and balanced interactive atlas packing
Gu, Floria Ziting
DOI : 10.14288/1.0450475
URI : http://hdl.handle.net/2429/92648 
Degree : Master of Science – MSc 
Graduation Date : 2025-11
Supervisor : Dr. Alla Sheffer 

A developer-centric compliance tool for serverless applications
Gupta, Praveen Kumar
DOI : 10.14288/1.0447498
URI : http://hdl.handle.net/2429/89871 
Degree : Master of Science – MSc 
Graduation Date : 2025-05
Supervisor : Dr. Aastha Mehta, Dr. Mohammad Shahrad

[no title]  
Hall, Jason Murray
Degree : Master of Science – MSc 
Graduation Date : 2025-05

Envisioning interventions to combat misinformation propagation on social media : insights from older adults’ approaches to credibility assessment and sharing decisions
Haque, Ishita
DOI : 10.14288/1.0447576
URI : http://hdl.handle.net/2429/89941 
Degree : Master of Science – MSc 
Graduation Date : 2025-05
Supervisor : Dr. Joanna McGrenere

Repurposing large pretrained diffusion models for unsupervised visual understanding and efficient adaptation
Hedlin, Eric
DOI : 10.14288/1.0450306
URI : http://hdl.handle.net/2429/92489 
Degree : Doctor of Philosophy – PhD 
Graduation Date : 2025-11
Supervisor : Dr. Kwang Moo Yi

Large pretrained text-conditioned image generation models learn a compositional and structured latent representation of visual concepts, showcasing their rich understanding of the world through their ability to generate diverse, coherent images. These models link text descriptions to visual concepts, unifying concepts across a range of conditions such as understanding the relationships between the text input and objects in a scene. This thesis explores how this link between text and visual concepts enables identifying consistent semantic regularities across images, where similar regions are mapped through the same text embedding. We show that this can be leveraged for tasks like semantic correspondence and estimating consistent keypoints, simply by optimizing the text embedding to activate highly in a specific region in the image for a given token. We also take advantage of the capacity of the model for one-shot personalization given only a single image. We leverage this by training hypernetworks to quickly estimate network weights for subject personalized generation, whose convergence is only possible due to the smooth underlying representation of concepts learned by these models. This PhD thesis leverages large pretrained diffusion models to address three key areas: semantic correspondence, unsupervised keypoint detection, and efficient hypernetwork-based adaptation for personalized model fine tuning. For semantic correspondence, we optimize text tokens to focus attention on specific regions in an image, leveraging the latent knowledge of large pretrained models to identify correspondences from a single image without additional supervision. For unsupervised keypoint detection, we localize text tokens across a collection of images to identify common keypoints, using a collection of images to focus the model on a specific concept, leveraging the knowledge within the pretrained model to generalize without ground truth keypoints. We also investigate hypernetwork-based methods for generating weights for large model personalization conditioned on a single image, providing an efficient alternative to compute intense optimization without requiring ground truth weights. This work highlights the versatility of diffusion models, extending their utility beyond image generation while proposing scalable, efficient solutions for downstream tasks of semantic correspondence, unsupervised keypoint estimation, and hypernetwork-based personalized model fine tuning.

Investigating fuzzing strategies in a CI/CD setup
Huang, Huicong (Madonna)
DOI : 10.14288/1.0447328
URI : http://hdl.handle.net/2429/89713 
Degree : Master of Science – MSc 
Graduation Date : 2025-05
Supervisor : Dr. Caroline Lemieux

Viability estimation for diffusion-based planning
Ioannidis, Nicholas
DOI : 10.14288/1.0448629
URI : http://hdl.handle.net/2429/90859 
Degree : Master of Science – MSc 
Graduation Date : 2025-05
Supervisor : Dr. Michiel van de Panne

Discourse-guided text-generation from knowledge graphs and image scene graphs
Ivanova, Inna
DOI : 10.14288/1.0449275
URI : http://hdl.handle.net/2429/91497
Degree : Master of Science – MSc 
Graduation Date : 2025-11
Supervisor : Dr. Giuseppe Carenini

Communcation-efficient algorithms for decentralized multi-task learning
Kuang, Yao
DOI : 10.14288/1.0450062
URI : http://hdl.handle.net/2429/92259
Degree : Master of Science – MSc 
Graduation Date : 2025-11
Supervisor : Dr. Michael Friedlander

Taking advantage of common assumptions in policy optimization and reinforcement learning
Lavington, Jonathan Wilder
DOI : 10.14288/1.0447187
URI : http://hdl.handle.net/2429/89582
Degree : Doctor of Philosophy – PhD 
Graduation Date : 2025-05 
Supervisor : Dr. Mark Schmidt, Dr. Frank Wood

This work considers training conditional probability distributions called policies, using simulated environments via gradient-based optimization methods. It begins by investigating the effects that complex model classes have on settings where a policy is learned through the imitation of expert data which is gathered through repeated environmental interaction. Next, it discusses how to build a gradient based optimizer which is tailored specifically to policy optimization where querying gradient information is expensive. We then consider policy optimization settings which contain imperfect expert demonstration, and design an algorithm which utilizes additional information available during training to improve the policy performance at test time and the efficiency of learning. Lastly, we consider how to generate behavioral data which satisfies hard constraints by using a combination of learned inference artifacts and a special variant of sequential Monte Carlo.

Towards realistic controllable driving simulators 
Lioutas, Vasileios
DOI : 10.14288/1.0448876
URI : http://hdl.handle.net/2429/91124
Degree : Doctor of Philosophy – PhD   
Graduation Date : 2025-11
Supervisor : Dr. Frank Wood

The development of autonomous vehicles requires extensive testing in simulated environments before deployment; however, current simulation approaches often fail to capture the complex, interactive nature of human driving behavior, relying instead on simplified or scripted agent behaviors. This thesis presents novel methods for generating realistic multi-agent driving behavior in simulated environments. We introduce ITRA (Imagining The Road Ahead), a generative model based on recurrent variational neural networks that captures complex spatial and temporal dependencies in driving behavior. To address ITRA's tendency to generate unsafe trajectories, we develop TITRATED, which combines amortized rejection sampling with differentiable infraction losses, and CriticSMC, a novel algorithm that enhances planning efficiency through learned value function heuristics. We then present Control-ITRA, enabling precise control over agent behavior through waypoint specification and target speeds while maintaining behavioral realism. Our extensive experimental evaluation demonstrates that these methods improve the realism and safety of simulated driving behavior while providing flexible control mechanisms for scenario generation, advancing the state-of-the-art in autonomous vehicle simulation and testing.

Social media clones : exploring the impact of social delegation with AI clones through a design workbook study
Liu, Jackie
DOI : 10.14288/1.0447712
URI : http://hdl.handle.net/2429/90083
Degree : Master of Science – MSc 
Graduation Date : 2025-05
Supervisor : Dr. Dongwook Yoon 

A bottom-up framework for cross-cultural evaluation of GPT-4o’s social norm biases via implicit narrative invocation
Liu, Zhuozhuo
DOI : 10.14288/1.0450081
URI : http://hdl.handle.net/2429/92274
Degree : Master of Science – MSc 
Graduation Date : 2025-11 
Supervisor : Vered Shwartz 

Indaleko : the unified personal index
Mason, William Anthony
DOI : 10.14288/1.0449905
URI : http://hdl.handle.net/2429/92101 
Degree : Doctor of Philosophy – PhD 
Graduation Date : 2025-11
Supervisor : Dr. Margo Seltzer, Dr. Ada Gavrilovska  

Personal information retrieval fails when systems ignore how human memory works. While existing platforms force keyword searches across isolated silos, humans naturally recall through episodic cues like when, where, and in what context information was encountered. This dissertation presents the Unified Personal Index (UPI), a memory-aligned architecture that bridges this fundamental gap. The Indaleko prototype demonstrates the UPI's feasibility on a 31-million file dataset spanning 160TB across eight storage platforms. By integrating temporal, spatial, and activity metadata into a unified graph database, Indaleko enables natural language queries like "photos near the conference venue last spring" that existing systems cannot process. The implementation achieves sub-second query responses through memory anchor indexing, eliminates cross-platform search fragmentation, and maintains perfect precision for well-specified memory patterns. Evaluation against commercial systems (Google Drive, OneDrive, Dropbox, Windows Search) reveals that all fail on memory-based queries, returning overwhelming result sets without contextual filtering. In contrast, Indaleko successfully processes multi-dimensional queries combining time, location, and activity patterns. The extensible architecture supports rapid integration of new data sources (10 minutes to 10 hours per provider) while preserving privacy through UUID-based semantic decoupling. The UPI's architectural synthesis bridges cognitive theory with distributed systems design, as demonstrated through the Indaleko prototype and rigorous evaluation. This work transforms personal information retrieval from keyword matching to memory-aligned finding, providing immediate benefits for existing data while establishing foundations for future context-aware systems.

What does the Adam optimizer actually adapt to?
Milligan, Alan
DOI : 10.14288/1.0450304 
URI : http://hdl.handle.net/2429/92483
Degree : Master of Science – MSc 
Graduation Date : 2025-11 
Supervisor : Dr. Mark Schmidt

Synthesizing device emulators
Noorafshan, Sepehr
DOI : 10.14288/1.0450324
URI : http://hdl.handle.net/2429/92499
Degree : Master of Science – MSc 
Graduation Date : 2025-11
Supervisor : Dr. Margo Seltzer

Physics-based character controllers with reinforcement learning
Reda, Daniele
DOI : 10.14288/1.0448165
URI : http://hdl.handle.net/2429/90456
Degree : Doctor of Philosophy – PhD 
Graduation Date : 2025-05
Supervisor : Dr. Michiel van de Panne

Physics-based character control has advanced significantly in recent years, with reinforcement learning (RL) emerging as a powerful method for producing general, versatile controllers. However, applying RL to humanoid control in animation and robotics poses fundamental challenges, including brittle policies, difficulties in exploring high-dimensional spaces, and a reliance on high-quality motion capture data. This thesis addresses these challenges, offering insights and methods in physics-based character control. We begin with a survey of learning methods for humanoid control, identifying the exploration problem as a fundamental limitation for developing general controllers. The absence of expert data and the complexity of state-action spaces hinder RL’s effectiveness. We review existing solutions to these challenges, highlighting their strengths and limitations. Next, we show the crucial role of environment design in RL. Poor design choices in key components, such as state representations, reward structures, or action spaces, can result in brittle and inefficient learning. Our analysis highlights how thoughtful environment enhances controller robustness and generalization. A key challenge in RL is the heavy reliance on extensive, high-quality motion capture datasets to guide learning. This raises the question: what happens if we lack data for specific motions, either because it is unavailable or unfeasible to collect? To address this, we introduce two strategies that reduce the dependency on motion capture data. The first approach uses a simplified physical model to provide motion priors, enabling controllers to learn complex behaviors without reference data. Through learning a brachiating controller, we show how the simplified model guides the center-of-mass trajectory and grasp timing, allowing the full model to efficiently learn swinging behaviors. The second strategy focuses on sparse motion retargeting, adapting data from unrelated characters to new morphologies. Our framework retargets sparse sensor data to physically simulated characters with different skeletal structures. Using physics as a prior to refine the kinematic motion and overcoming the exploration problem, our method produces robust real-time controllers for applications in virtual reality (VR) and robotics. Together, these contributions enable the development of robust, adaptable controllers, broadening RL’s applicability to physics-based animations and advancing the state of the art in learning character control.

Learning dynamics of deep learning -- force analysis of deep neural networks
Ren, Yi (Joshua)
DOI : 10.14288/1.0450240
URI : http://hdl.handle.net/2429/92432
Degree : Doctor of Philosophy – PhD 
Graduation Date : 2025-11  
Supervisor : Dr. Danica Sutherland

This thesis investigates the learning dynamics of deep learning systems through a local, physics-inspired analytical lens. Motivated by the need for fine-grained insights into model behavior, we begin with the step-wise influence that a single training example exerts on a specific observing example during learning. Central to our approach is the proposed AKG decomposition, which dissects this influence into three interpretable components: similarity (K), normalization (A), and prediction gap (G). This decomposition enables an analogy with classical force analysis: the force originates from G, is shaped by AK, and is ultimately applied to the target object, e.g., to the model confidence, output, hidden representations, or parameters. Building upon this foundation, we gradually scale the analysis from individual interactions to cumulative effects over time, akin to tracking an object’s motion under multiple forces. We apply it to the following problems. Supervised classification: We study the learning trajectories of examples with varying difficulty and reveal an interesting "zig-zag" pattern that emerges during optimization. Our analysis explains this behavior and inspires a novel knowledge distillation method, Filter-KD, which improves supervision signals for student models. Large language model (LLM) finetuning: We extend the framework to account for the autoregressive nature of LLMs and the presence of negative gradients. The unified perspective explains behaviors across finetuning methods such as SFT, DPO, and GRPO. We also highlight the critical role of negative gradients. In particular, we identify the "squeezing effect": a counterintuitive phenomenon caused by improperly applied gradient ascent. Representation learning: We explore the dynamics of hidden features, revealing how adaptation energy and directions influence the feature drift. Our analysis leads to a provable pattern of feature adaptation in a head-probing then finetuning pipeline, offering insights and inspiring several practical strategies. Simplicity bias and compositional learning: Revisiting foundational questions about why structured representations are learned faster, we apply our framework to a compositional learning setting. Our findings align with principles such as Occam’s Razor and the idea of "compression for AGI," offering a novel dynamical explanation rooted in compression and learning speed.

Indirect measurement techniques for detecting and evading QUIC censorship
Sengottuvelavan, Karthik Nishanth
DOI : 10.14288/1.0450509 
URI : http://hdl.handle.net/2429/92686
Degree : Master of Science – MSc 
Graduation Date : 2025-11
Supervisor : Nguyen Phong Hoang

Talking to an AI mirror : designing self-clone chatbots for enhanced engagement in digital mental health support
Shirvani, Mehrnoosh Sadat
DOI : 10.14288/1.0448512
URI : http://hdl.handle.net/2429/90722
Degree : Master of Science - MSc
Graduation Date : 2025-05
Supervisor : Dr. Dongwook Yoon

Workload-aware SQL query recommendation using retrieval-augmented generation
Soltan Aghai, Ehsan
DOI : 10.14288/1.0448564
URI : http://hdl.handle.net/2429/90790
Degree : Master of Science – MSc 
Graduation Date : 2025-05
Supervisor : Dr. Rachel Pottinger

Optimization with explorable uncertainty
Tabatabaee, Seyed Ali
DOI : 10.14288/1.0449643
URI : http://hdl.handle.net/2429/91837
Degree : Doctor of Philosophy – PhD 
Graduation Date : 2025-11
Supervisor : Dr. Will Evans

Many real-world problems involve elements (e.g., clients, jobs, etc.) with uncertain properties. Acquiring more accurate properties of these elements is often costly but sometimes necessary for high-quality solutions. The goal is to find such solutions through cost-effective strategies for obtaining more accurate information. The described model is often referred to as explorable uncertainty. This thesis studies optimization problems from the domains of facility location and job scheduling in the explorable uncertainty model. First, we study center problems with moving entities of bounded speed in Euclidean space, where the movement of the entities is unpredictable and processing must be done in real-time. Center problems involve determining the location of a facility to serve a set of entities while optimizing a specified objective function. In particular, we investigate computing the 1-center, centroid, center of mass, and 1-median for a set of moving entities. Next, motivated by the connections observed between these problems and perpetual scheduling problems, we shift our focus to study the latter in more depth. Perpetual scheduling problems involve jobs that require recurring processing. The goal is to process these jobs while optimizing a specified objective function. More specifically, we investigate two prominent examples of perpetual scheduling problems, namely the bamboo trimming problem and the windows scheduling problem, in settings that have not been considered before. Finally, we study these two scheduling problems in the model of explorable uncertainty, where jobs’ processing requirements can be reduced by taking some actions. We provide novel algorithms for the problems considered throughout this thesis. Our results contribute to a better understanding of the various forms of explorable uncertainty and the additional intricacy introduced to optimization problems when considered in this model.

IRBlock : a large-scale measurement study of the Great Firewall of Iran
Tai, Jonas
DOI : 10.14288/1.0449910
URI : http://hdl.handle.net/2429/92118
Degree : Master of Science - MSc
Graduation Date : 2025-11 
Supervisor : Dr. Nguyen Phong Hoang

Agent persona design for engagement in virtual dialogic learning environments
Tanprasert, Thitaree
DOI : 10.14288/1.0448341
URI : http://hdl.handle.net/2429/90635
Degree : Doctor of Philosophy – PhD 
Graduation Date : 2025-05
Supervisor : Dr. Dongwook Yoon

Asynchronous online learning (AOL) (e.g., online courses, video-sharing platforms) has become popular for its variety, accessibility, and flexibility. However, AOL often lacks the rich social interactions of traditional classrooms, lowering learners' engagement and, consequently, their evaluative performance. The recent advancement of Large-language Models (LLMs) illuminates a potential solution of using AI agents for real-time dialogic learning---a pedagogical approach involving dialogues, which has been shown to enhance learners' behavioral, emotional, and cognitive engagement. This approach is closely intertwined with the characteristics and perception of the interactors (learners, educators, etc.), so the design of the agent's persona (behaviors, appearances, and identity cues) is crucial. However, although many persona attributes can be simulated with AI, designing agents' persona presents two challenges: (1) humans respond to the same trait in human and AI agents differently, so existing frameworks of human-human interactions cannot be applied directly; and (2) LLM-based agent's behaviors may fluctuate due to learners' input, necessitating a new conceptualization of behaviors in empirical studies. In this thesis, I aim to address these challenges and show that persona design of educational agents in dialogic learning environments improves learning engagement for asynchronous, online learners. I demonstrate this via three research projects, which cover three social learning experiences. The first project focuses on debates (learners argue with agents with the opposite stance from them) and how the agent's social identity and rhetorical styles impact their influence. The second project concerns collaborative activities (learners work with peer agents with different values from theirs towards the same goal) and the impacts of the agent's collaborative strategies and the disclosure of their strategies to learners. The third project is about vicarious dialogues (learners observe conversations between multiple characters) and the design of virtual group dynamics. The findings demonstrate that theoretically grounded and learner-centric persona designs of agents improve learning engagement. They also highlight the importance of contextual factors (e.g., activity types, learners' values) when applying social learning theories to agents, generate design implications for agent persona design beyond AOL contexts, and raise ethical and pedagogical considerations for the upcoming era of AI-powered education.

Graph-augmented deep learning using literature-informed biological priors for predicting perturbations in single-cell RNA sequencing
Tu, Lin Shuan (Wilson)
DOI : 10.14288/1.0449273
URI : http://hdl.handle.net/2429/91500
Degree : Master of Science - MSc
Graduation Date : 2025-11  
Supervisor : Dr. Jiarui Ding, Dr. Alexander Wyatt 

Hiring under uncertainty and competition : a work on various extensions of the secretary problem
Turkieltaub Melo, Abner
DOI : 10.14288/1.0449765
URI : http://hdl.handle.net/2429/91958
Degree : Doctor of Philosophy – PhD 
Graduation Date : 2025-11
Supervisor : Dr. Hu Fu, Dr. Bruce Shepherd

In this dissertation, we study different online hiring problems related to the Secretary Problem. In the Secretary Problem, an employer sees a sequence of candidates. Each time a new candidate arrives, the employer makes an irrevocable choice on whether to hire based only on the relative ranking of the candidates seen so far. The employer tries to maximize the probability of hiring the best. It is known that the optimal strategy hires the best with probability 1/e. We study several extensions. For many of them, we work on an infinite arrival regime. This allows us to apply a lemma we prove characterizing the number of “promising candidates” in any given time interval. For a single employer trying to hire k candidates, we produce a tight analysis of some simple strategies. We also study in detail the case k = 2. We derive an optimal algorithm under the objective known as “probability-competitiveness”. We also study the case in which the two selected candidates must be independent in a given matroid. For multiple employers we consider three models. First, for employers seeing the same sequence of candidates, we compute Nash Equilibria for two and three employers. We also derive general properties of the Nash Equilibria for any number of employers. For employers seeing the same set of candidates, but in different order, we compute a Nash Equilibria for two employers. Finally, we consider employers with different sets of candidates, competing for whom hires first while trying to get their best candidate. As motivation, imagine different research groups within the same department trying to hire someone in their area. We show that without any regulation, competition pushes employers to hire too early, making extremely unlikely the hiring of a top candidate. We also compute the social optimum and propose different regulations that incentivize employers to align with the social optimum. Finally, we study oblivious Online Contention Resolution Schemes (OCRSs) for matroids. This problem has an interesting link to the Matroid Secretary Conjecture. We provide an optimal 1/e-selectable oblivious OCRS for selecting a single item. We also prove non-existence of constant-selectable oblivious OCRSs for general matroids.

Side-channel security in networks : from the internet to interconnects
Vora, Rut
DOI : 10.14288/1.0448228
URI : http://hdl.handle.net/2429/90513
Degree : Master of Science – MSc 
Graduation Date : 2025-05
Supervisor : Dr. Aastha Mehta

Cache side-channel attacks on language runtimes
Wang, Yayu
DOI : 10.14288/1.0448547
URI : http://hdl.handle.net/2429/90781
Degree : Master of Science – MSc 
Graduation Date : 2025-05
Supervisor : Aastha Mehta

Structured amortized variational inference
Weilbach, Christian
DOI : 10.14288/1.0449997
URI : http://hdl.handle.net/2429/92167
Degree : Doctor of Philosophy - PhD
Graduation Date : 2025-11  
Supervisor : Dr. Frank Wood

This thesis explores how structural knowledge about inference problems can be automatically mapped into solution mechanisms or inference artifacts, significantly enhancing efficiency and scalability. Initially, programming language theory is employed to implement a compiler that extracts structural knowledge from problem specifications and transforms it into a faithful inverse graph for inference. This graph constrains a neural network within a continuous normalizing flow (CNF) framework, providing efficient, high-quality inference for problems with a limited number of variables. However, the CNF framework faces inherent scalability and efficiency limitations. To address these, the approach is refined by structuring the attention mechanism in more powerful transformer neural networks within simulation-free denoising diffusion probabilistic models (DDPMs). The enhanced graphically structured diffusion model (GSDM) framework effectively handles various algorithmic problems, such as matrix factorization, Sudoku solving, and sorting. These tasks are integral subproblems in larger scientific inference settings. Furthermore, the framework is extended to scale beyond GPU memory limitations and integrated with the simulation-based inference community, facilitating its direct application in simulators.

Closures under span and intersection of three and four subspaces
Ye, Xinyi
DOI : 10.14288/1.0447515
URI : http://hdl.handle.net/2429/89887
Degree : Master of Science – MSc 
Graduation Date : 2025-05
Supervisor : Dr. Joel Friedman