Computer Science

Permanent URI for this collectionhttps://uwspace.uwaterloo.ca/handle/10012/9930

This is the collection for the University of Waterloo's Cheriton School of Computer Science.

Research outputs are organized by type (eg. Master Thesis, Article, Conference Paper).

Waterloo faculty, students, and staff can contact us or visit the UWSpace guide to learn more about depositing their research.

Browse

Now showing 1 - 20 of 1718

Polymorphic Type Qualifiers
(University of Waterloo, 2025-10-21) Edward, Lee
Type qualifiers offer a simple yet effective mechanism for extending existing type systems to deal with additional constraints or safety requirements. For example, the const qualifier is a popular mechanism for annotating existing types to signify that the value in question is read-only in addition. A variable of type const int is both an integer and also cannot be written to. While type qualifiers themselves are well-studied, polymorphism over type qualifiers remains an area less well examined. This has led to a number of ill-desired outcomes. For one, many practical systems implementing type qualifiers in their type systems simply ignore their interaction with generic types. Other systems implement polymorphism with seemingly unique and ad-hoc rules for dealing with qualifiers. In this thesis, we show that this does not need to be the case. We start by examining three well-known qualifier systems: systems for tracking immutability, function colour, and captured variables, and show that despite their differences that they share surprising common structure. We then give a design recipe, inspired by that structure, using the mathematical structure of free lattices for modelling polymorphism over type qualifiers, to give a framework for polymorphic type qualifiers. We then show that our design recipe precisely captures this structure by recasting those three existing systems in our framework for qualifier polymorphism by free lattices. Finally, we extend type qualifiers from ranging over lattices to type qualifiers ranging over Boolean algebras, which we then use to extend an existing effect system with effect exclusion to support subeffecting as well via subqualification and subtyping.
ReSlide: Towards Effective Presentation Authoring for Evolving Narratives and Contextual Constraints
(University of Waterloo, 2025-09-29) Zeng, Linxiu
Authoring slides for a public presentation requires speakers to navigate multiple contextual constraints that directly shape how content is structured and delivered. We first conducted a formative study with ten experienced presenters to understand what constraints are prioritized and how they influence slide deck creation. We identified three key constraints---time, audience, and communicative intent---and challenges in integrating them into slide authoring for a single presentation and long-term needs with diverse narratives. We designed ReSlide, a presentation authoring tool that helps presenters create and reuse slides by bridging contextual constraints with evolving narratives. We evaluated ReSlide in a within-subjects study with 12 participants against a baseline tool, and in an exploratory study with eight professional presenters using only ReSlide. Results indicate that Reslide's novel features of constraint awareness and multi-granular slide reuse helped presenters effectively craft presentations in both one and multiple authoring cycles.
Concert: Elementary VR Interactions Augmented By Bimanual Input
(University of Waterloo, 2025-09-25) Zhang, Futian
Many elementary tasks in Virtual Reality (VR)—such as pointing, locomotion, and command selection—are still predominantly performed through unimanual interaction, even though the non-dominant hand often remains available. When these tasks require multiple sequential steps, their frequent use can increase overall interaction time and user effort. By engaging both hands in complementary roles, some of these tasks can be streamlined into more efficient, parallelized actions without sacrificing intuitiveness. This thesis, Concert, investigates how bimanual input can optimize such underexplored opportunities through three novel interaction techniques. The first, Conductor, is an intersection-based 3D pointing method where the dominant hand controls an aiming ray while the non-dominant hand positions an intersecting plane to determine cursor depth. Compared to Raycursor, a state-of-the-art VR pointing technique, Conductor enables faster and more accurate target selection. The second project, Fly The Moon To Me (Locomoontion), extends Conductor to 3D object manipulation for locomotion, integrating vertical (height) and horizontal (planar) adjustments into a single fluid operation. Users generate a preview copy of a target object, reposition it precisely in 3D space, and then align the original object and surrounding environment to the preview. In a teleportation task, Locomoontion outperformed Point & Teleport and Point & Tug, reducing completion time and physical effort. The third project, Drum Menu, accelerates command selection through bimanual shortcuts inspired by marking menus. Users select commands by joystick rotation, stroke drawing, or directional pointing, with the bimanual variant granting simultaneous access to two menu levels. In user studies, bimanual configurations were faster than unimanual ones for a 4-item layout, with participants preferring the bimanual joystick menu; however, 8-item layouts increased error rates among experienced users. Collectively, these studies show that many common VR tasks can be made more efficient through carefully designed bimanual interaction, providing concrete design guidelines for integrating both hands into future immersive systems.
Conjunctive Queries with Negations: Bridging Theory and Practice
(University of Waterloo, 2025-09-24) Li, Boyi
Antijoin, given its significant expressive power, has numerous applications in relational data analytics. Notwithstanding its importance, there remains great research potential in antijoin processing. In practical database systems, existing techniques to process antijoins are still considered rudimentary, building upon heuristics and cost-based optimization strategies that offer no theoretical guarantees. Meanwhile, the database theory community has proposed algorithms for antijoins with strong theoretical guarantees, yet these algorithms build upon specialized, complicated data structures and have not made their way to practice. In light of such gap between theory and practice, we propose new algorithms for antijoin processing in this thesis. Not only do our new algorithms provide the same theoretical guarantees as the state-of-the-art algorithm, but they also use only basic relational operations. The latter property enables our new algorithms to be rewritten in basic SQL statements, allowing an easy, system-agnostic integration into any SQL-based database system. We then empirically evaluate one of our new algorithms, rewritten in SQL, over real-life graph datasets with a variety of SQL database systems. Experimental results show order-of-magnitude improvements of our new algorithm over vanilla SQL queries.
Oblivious Multi-Way Band Joins: An Efficient Algorithm for Secure Range Queries
(University of Waterloo, 2025-09-22) Wei, Ruidi
This thesis introduces the first efficient oblivious algorithm for acyclic multi-way joins with band conditions, extending the classical Yannakakis algorithm to support inequality predicates (>, <, ≥, ≤) without leaking sensitive information through memory access patterns. Band joins, which match tuples over value ranges rather than exact keys, are widely used in temporal, spatial, and proximity-based analytics but present challenges in oblivious computation. Our approach employs a dual-entry technique that transforms range matching into cumulative sum computations, enabling multiplicity computation in an oblivious manner. The algorithm achieves O(N log N + k · OUT log OUT) complexity, where k is the number of tables in the join query, N is the input size, and OUT is the output size, matching state-of-the-art oblivious equality joins up to a factor of k while supporting full band constraints. We implement the method using Intel SGX with batch processing and evaluate it on the TPC-H benchmark dataset, demonstrating practical performance and strong obliviousness guarantees under an honest-but-curious adversary model.
Toward General-Purpose Monte Carlo PDE Solvers for Graphics Applications
(University of Waterloo, 2025-09-22) Sugimoto, Ryusuke
This thesis develops novel Monte Carlo methods for solving a wide range of partial differential equations (PDEs) relevant to computer graphics. While traditional discretization-based approaches efficiently compute global solutions, they often require expensive global solves even when only local evaluations are needed, and can struggle with complex or fine-scale geometries. Monte Carlo methods based on the classical Walk on Spheres (WoS) approach [Muller 1956] offer pointwise evaluation with strong geometric robustness, but in practice, their application has been largely limited to interior Dirichlet problems in volumetric domains. We significantly broaden this scope by designing versatile Monte Carlo solvers that handle a diverse set of PDEs and boundary conditions, validated through comprehensive experimental results. First, we introduce the Walk on Boundary (WoB) method [Sabelfeld 1982, 1991] to graphics. While retaining WoS’s advantages, WoB applies to a broader range of second-order linear elliptic and parabolic PDE problems: various boundary conditions (Dirichlet, Neumann, Robin, and mixed) in both interior and exterior domains. Because WoB is based on boundary integral formulations, its structure more closely parallels Monte Carlo rendering than WoS, enabling the application of advanced variance reduction techniques. We present WoB formulations for elliptic Laplace and Poisson equations, time-dependent diffusion problems, and develop a WoB solver for vector-valued Stokes equations. Throughout, we discuss how sampling and variance reduction methods from rendering can be adapted to WoB. Next, we address the nonlinear Navier-Stokes equations for fluid simulation, whose complexity challenges Monte Carlo techniques. Employing operator splitting, we separate nonlinear terms and solve the remaining linear terms with pointwise Monte Carlo solvers. Recursively applying these solvers with timestepping yields a spatial-discretization-free method. To deal with the resulting exponential computational cost, we also propose cache-based alternatives. Both vorticity- and velocity-based formulations are explored, retaining the advantages of Monte Carlo methods, including geometric robustness and variance reduction, while integrating traditional fluid simulation techniques. We then propose Projected Walk on Spheres (PWoS), a novel solver for surface PDEs, inspired by the Closest Point Method. PWoS modifies WoS by projecting random walks onto the surface manifold at each step, preserving geometric flexibility and discretization-free, pointwise evaluation. We also adapt a noise filtering technique for WoS to improve PWoS. Finally, we outline promising future research directions for Monte Carlo PDE solvers in graphics, including concrete proposals to enhance WoB.
Diffusion-Based Generative Modeling of Financial Time Series
(University of Waterloo, 2025-09-22) Briazkalo, Mykhailo
Financial time series demonstrate complex stochastic dynamics such as volatility clustering, heavy tails, and sudden jumps that are difficult to capture with traditional parametric models. Deep generative models offer a flexible alternative for learning unknown data distributions, but their application to financial data remains limited. In this thesis, we propose a diffusion-based generative framework for modeling financial time series. Building on the Elucidated Diffusion Model, originally developed for image synthesis, we adapt its architecture to multivariate sequential data and integrate the Ambient Diffusion framework as a variance correction mechanism. We provide a theoretical analysis connecting excess variance in standardized outputs with volatility bias and derive an analytical rule for selecting the ambient noise level. We evaluate the proposed approach using a comprehensive framework that combines statistical similarity, parameter recovery, option pricing, and risk metrics. Across synthetic models (GBM, Heston, Merton) and real-world datasets (SPY, AAPL, NVDA, BTC), our Ambient Diffusion-based method consistently improves distributional alignment, volatility recovery, and option pricing over the EDM baseline, highlighting its potential for quantitative modeling, scenario generation, and risk management.
Fine-Grained Visual Entity Linking through Promptable Segmentation: Applications in Medical Imaging
(University of Waterloo, 2025-09-19) Carbone, Kathryn
Image analysis in domains that produce large amounts of complex visual data, like medicine, is challenging due to time and labour-constraints on domain experts. Visual entity linking (VEL) is a preliminary image processing task which links regions of interest (RoIs) to known entities in structured knowledge bases (KBs), thereby using knowledge to scaffold image understanding. We study a targeted VEL problem in which a specific user-highlighted RoI within the image is used to query a textual KB for information about the RoI, which can support downstream tasks such as similar case retrieval and question answering. For example, a doctor reviewing an MRI scan may wish to obtain images with similar presentations of a medically relevant RoI, such as a brain tumor, for comparison. By linking this RoI to its corresponding KB document, search of an imaging database with VEL-guided automatically-generated tags can be performed in a knowledge-aware manner based on exact or semantically similar entity tag matching. Cross-modal embedding models like CLIP present straightforward solutions through the dual encoding of KB entries and either whole images or cropped RoIs, which can then be matched by a vector similarity search between these respective learned representations. However, using the whole image as the query may retrieve KB entries related to other aspects of the image besides the RoI; at the same time, using the RoI alone as the query ignores context, which is critical for recognizing and linking complex entities such as those found in medical images. To address these shortcomings, this thesis proposes VELCRO—visual entity linking with contrastive RoI alignment—which adapts an image segmentation model to VEL using contrastive learning by aligning the contextual embeddings produced by its decoder with the KB. This strategy preserves the information contained in the surrounding image while focusing KB alignment specifically on the RoI. To accomplish this, VELCRO performs segmentation and contrastive alignment in one end-to-end model via a novel loss function that combines the two objectives. Experimental results on medical VEL show that VELCRO achieves an overall linking accuracy of 95.2% compared to 83.9% for baseline approaches.
Exploiting Zero-Entropy Data for Efficient Deduplication
(University of Waterloo, 2025-09-15) Al Jarah, Mu'men
As the volume of digital data continues to grow rapidly, efficient data reduction techniques, such as deduplication, are essential for managing storage and bandwidth. A key step in deduplication is file chunking, which is typically performed using Content-Defined Chunking (CDC) algorithms. While these algorithms have been studied under random data, their performance in the presence of zero-entropy data, where long sequences of identical bytes appear, has not been explored. Such zero-entropy data are common in real-world datasets and introduce challenges for CDC in deduplication systems. This thesis studies the impact of zero-entropy data on the performance of both hash-based and hashless state-of-the-art CDC algorithms. The results show that existing algorithms, particularly hash-based ones, are inefficient at detecting and handling zero-entropy blocks, especially when these blocks are small, which reduces space savings. To address this issue, I propose ZERO (Zero-Entropy Region Optimization), a system that can be integrated into the deduplication pipeline. ZERO identifies and extracts zero-entropy blocks prior to chunking, compresses them using Run-Length Encoding (RLE), and stores their metadata for later reconstruction. ZERO improves deduplication space savings by up to 29% without impacting throughput.
Precise and Scalable Constraint-Based Type Inference for Incomplete Java Code Snippets in the Age of Large Language Models
(University of Waterloo, 2025-09-08) Dong, Yiwen
Online code snippets are prevalent and are useful for developers. These snippets are commonly shared on websites such as Stack Overflow to illustrate programming concepts. However, these code snippets are frequently incomplete. In Java code snippets, type references are typically expressed using simple names, which can be ambiguous. Identifying the exact types used requires fully qualified names typically provided in import statements. Despite their importance, such import statements are only available in 6.88% of Java code snippets on Stack Overflow. To address this challenge, this thesis explores constraint-based type inference to recover missing type information. It also proposes a dataset for evaluating the performance of type inference techniques on Java code snippets, particularly large language models (LLMs). In addition, the scalability of the initial inference technique is improved to enhance applicability in real-world scenarios. The first study introduces SnR, a constraint-based type inference technique to automatically infer the exact type used in code snippets and the libraries containing the inferred types, to compile and therefore reuse the code snippets. Initially, SnR builds a knowledge base of APIs, i.e., various facts about the available APIs, from a corpus of Java libraries. Given a code snippet with missing import statements, SnR automatically extracts typing constraints from the snippet, solves the constraints against the knowledge base, and returns a set of APIs that satisfies the constraints to be imported into the snippet. When evaluated on the StatType-SO benchmark suite, which includes 267 Stack Overflow code snippets, SnR significantly outperforms the state-of-the-art tool Coster. SnR correctly infers 91.0% of the import statements, which makes 73.8% of the snippets compilable, compared to Coster’s 36.0% and 9.0%, respectively. The second study evaluates type inference techniques, particularly of LLMs. Although LLMs demonstrate strong performance on the StatType-SO benchmark, the dataset has been publicly available on GitHub since 2017. If LLMs were trained on StatType-SO, then their performance may not reflect how the model would perform on novel, real-world code, but rather result from recalling examples seen during training. To address this, this thesis introduces ThaliaType, a new, previously unreleased dataset containing 300 Java code snippets. Results reveal that LLMs exhibit a significant drop in performance when generalizing to unseen code snippets, with up to 59% decrease in precision and up to 72% decrease in recall. To further investigate the limitations of LLMs in understanding the execution semantics of the code, semantic-preserving code transformations were developed. Analysis showed that LLMs performed significantly worse on code snippets that are syntactically different but semantically equivalent. Experiments suggest that the strong performance of LLMs in prior evaluations was likely influenced by data leakage in the benchmarks, rather than a genuine understanding of the semantics of code snippets. The third study enhances the scalability of constraint-based type inference by introducing Scitix. Constraint-solving becomes computationally expensive using a large knowledge base in the presence of unknown types (e.g. user-defined types) in code snippets. To improve scalability, Scitix represents certain unknown types as Any, ignoring such types during constraint solving. Then an iterative constraint-solving approach saves on computation and skips constraints involving unknown types. Extensive evaluations show that the insights improve both performance and scalability compared to SnR. Specifically, Scitix achieves F1-scores of 96.6% and 88.7% on StatType-SO and ThaliaType, respectively, using a large knowledge base of over 3,000 jars. In contrast, SnR consistently times out, yielding F1-scores close to 0%. Even with the smallest knowledge base, where SnR does not time out, Scitix reduces the number of errors by 79% and 37% compared to SnR. Furthermore, even with the largest knowledge base, Scitix reduces error rates by 20% and 78% compared to state-of-the-art LLMs. This thesis demonstrates the use of constraint-based type inference for Java code snippets. The proposed approach is evaluated through a comprehensive analysis that contextualizes its performance in the current landscape dominated by LLMs. The ensuing system, Scitix, is both precise and scalable, enhancing the reusability of Java code snippets.
Combining High-Level-Synthesis and Register-Transfer Level Design for Programmable Hardware
(University of Waterloo, 2025-09-08) Mohammadtaheri, Kimiya
As network speeds continue to grow to hundreds of Gbps and beyond, offloading transport-layer functionality to programmable Network Interface Cards (NICs) has gained traction for improving performance and enabling higher-level offloads. FPGAs on NICs offer low latency and high throughput but are notoriously difficult to program. To reduce developer effort for hardware transport, prior works propose hardware architectures with reusable fixed-function modules for common transport data structures and operations, and programmable modules for protocol-specific operations. However, they either still require intricate Verilog programming for their programmable modules or are tied to specific protocols or pipeline abstractions. In this work, we explore the design of a programmable hardware architecture for transport, called HTraP, that (1) offers a simple, intuitive programming interface without requiring detailed understanding of the rest of the architecture for effective programming, and (2) supports a broad range of transport protocols with varying levels of complexity. HTrap has one main programmable module that can be programmed to capture the core protocol logic in simple C code that is amenable to automated, efficient hardware generation with HLS. The generated hardware can then plug into the rest of HTraP which implements complex transport operations through a protocol-agnostic interface.
Instance Segmentation with Occlusion Order Supervision: Two Problems
(University of Waterloo, 2025-09-05) Yu, Cheuk-To Jeremy
Joint architectures for predicting instance masks with additional depth-related information have shown improvements in performance on both the original segmentation tasks and their corresponding depth-related tasks. Closely related to depth, occlusion can also provide strong cues for segmentation. Although occlusion ordering between instances provides less supervision than full pixel-wise depth, occlusions can be easily annotated for existing datasets. Motivated by the above, we propose two problems that incorporate occlusion information into standard segmentation tasks with their corresponding methods. For both methods, we explore the level of supervision occlusion provides in terms of segmentation accuracy by comparing our results with baseline segmentation approaches trained without occlusion supervision. Firstly, we develop an end-to-end framework to perform instance segmentation and per-instance global occlusion order prediction simultaneously by appending a global occlusion order head to standard instance segmentation architectures. Our approach to occlusions differs from most prior work. Prior work performs occlusion estimation in a local pairwise manner: given two instances, classify one of them as the occluder. Due to locality, for scenes known not to have occlusion cycles, such approaches can (and do) produce occlusion cycles, which are errors. Unlike most prior work, we directly label instances with their occlusion-order labels, where an instance with a larger label occludes any neighboring instances with smaller labels. This approach is cycle-free by design. Using cross-entropy with occlusion-order labels fails as occlusion-order labels do not have a fixed semantic meaning. Therefore, we develop a novel regularized loss function for successful training. Our framework achieves high occlusion-order accuracy with improved performance in instance segmentation, likely due to the added supervision. Secondly, we develop a new direction for occlusion-supervised amodal instance segmentation (AIS). AIS is an emerging task that segments the complete object instance, both the visible and occluded parts. All prior work for AIS can be divided into two groups: (i) methods constructing a synthetic amodal dataset; (ii) methods using human-annotated amodal masks. The drawback of methods in the first group is that constructing a realistic synthetic dataset is difficult. The drawback of methods in the second group is that human annotation is prone to error, as humans must reason about invisible regions. Our method requires neither a synthetic dataset nor human-annotated amodal masks, but instead uses ground-truth modal masks and the pairwise occlusion order between instances. By using the occlusion order to reason about the visible and invisible parts of the amodal mask, we develop effective loss functions for the visible and occluded parts of the predicted mask. We achieve comparable accuracies to those of the same architecture trained on human-annotated amodal masks, despite not using amodal masks for training.
Realizing Privacy Design Strategies for Back-end Web Architectures
(University of Waterloo, 2025-09-03) Mazmudar, Miti
In today's big data world, companies collect extensive profiles of users' personal information. This kind of extensive data collection risks exposing users' personal information in larger, more severe and frequent data breaches. Furthermore, software development practices lead developers to processing this sensitive information for purposes orthogonal to the core functionality offered by the business. Users' sensitive data has often been shared with business partners and other third parties in violation of the business' privacy policy, and in a breach of the users' trust. Hoepman proposes eight privacy design strategies, in order to guide software developers to incorporate privacy as a central element within the software development process. These privacy design strategies underpin the design of, and draw examples from, various privacy-enhancing technologies (PETs) that are deployed today, including differential privacy, anonymity networks and so on. Nevertheless, software engineers and privacy engineers have found that realizing these privacy design strategies within the back-end web architectures that they maintain remains challenging. Some privacy design strategies, such as the Inform and Control strategies, are commonly realized, by informing users through privacy policies and equipping them with controls. However, the remaining strategies are not commonly translated into practice. We consider different back-end web architectures, including both centralized and distributed web architectures, and each contribution in this thesis operationalizes one or more privacy design strategies for the respective back-end architecture. In our first system, namely CacheDP, we examine relational DBMSes, and reduce the granularity of sensitive data revealed to data analysts using these systems, thereby realizing the Abstract strategy. In our second and third systems, namely DHTPIR and Peer2PIR, we delve into peer-to-peer networks and distributed hash tables. Both of these contributions seek to minimize the data revealed to other peers while a peer queries for, and retrieves, files from the network, and thus, they realize the Minimize strategy. However, both systems do so differently; DHTPIR separates the sensitive content of the query across multiple peers in the network, thereby realizing the Separate strategy, while Peer2PIR hides the content of the query using encryption, and thus operationalizes the Hide strategy. In our final system, namely Prose, we focus on a distributed microservice architecture, conceptualizing technical requirements that cater towards the needs of privacy engineers and software engineers, and implements the Enforce and Demonstrate strategies for this architecture. Considering the Minimize strategy as an example, we distinguish betweeen PETs that realize a privacy design strategy once within the underlying architecture, without further interventions as the architecture evolves, as is the case with DHTPIR and Peer2PIR, with PETs that realize a strategy in a way such that the PET can respond to changes in the architecture, as in Prose. We thus relate our privacy-enhancing technologies to the privacy design strategies that they realize, highlighting combinations of privacy design strategies that are realized simultaneously, and illustrating how some privacy design strategies may be used to indirectly realize others. Importantly, our work shows that PETs can be designed to systematically and continually realize privacy design strategies, such as the Enforce, Demonstrate and Minimize strategies. Finally, we evaluate our privacy-enhancing technologies and show that they are effective, in terms of continuing to providing the desired functionality but with the added privacy guarantee afforded by the privacy design strategy, and efficient, in terms of low communication and computation overheads over the original architecture.
Efficient Algorithm with No-Regret Bound for Sleeping Expert Problem
(University of Waterloo, 2025-08-29) Lin, Junhao
The sleeping experts problem is a variant of decision-theoretic online learning (DTOL) where the set of available experts may change over time. In this thesis, we study a special case of the sleeping experts problem with constraints on how the set of available experts can change. The benchmark we use is ranking regret, which is a common benchmark used in sleeping experts problem. Previous research shows that achieving sub-linear ranking regret bound in the general sleeping experts problem is NP-hard, so we relax the sleeping experts problem by imposing constraints on how the set of available experts may change. Under those constraints, we present an efficient algorithm which achieves a sub-linear ranking regret bound.
Anonymous Communication within Highly Repressive Regions
(University of Waterloo, 2025-08-28) Kamali, Sina
Repressive governments are increasingly resorting to restricting the flow of information during times of political unrest or civil resistance. In recent years, this restriction has taken new forms, expanding beyond traditional censorship tactics such as IP and DNS blocking to the use of Intranets (internal networks) and complete internet blackouts. To address this, many tools have been proposed to support users during such bleak times. However, most of these tools either overlook the importance of user anonymity or underestimate the lengths to which a censor might go to de-anonymize users. In this thesis, we propose two new tools to help users reclaim their freedom and privacy under these restrictive conditions. First, we present Anix, an anonymous, blackout-resistant microblogging app that supports message endorsing, allowing users to avoid falling victim to misinformation, even without internet access. Second, we introduce Huma, a censorship circumvention tool that operates via web protocol tunneling and leverages deferred traffic replacement to counter a wide range of adversarial attacks. Additionally, Huma can function as a blocking-resistant anonymous communication system within Intranets. Together, these tools offer secure and private solutions for users facing the aforementioned restrictive scenarios. We implement and evaluate both tools to demonstrate their practicality and release our code as an open-source contribution to the community.
Rotors for Moving Frames in Projective Geometric Algebra
(University of Waterloo, 2025-08-28) Leger, Zachary Charles Joseph
Projective Geometric Algebra (PGA) is a relatively new mathematical model to describe rigid-body motions. As such, there has yet to be in-depth research on using PGA in conjunction with differential geometry. In particular, PGA's ability to describe rigid-body motions appears to be well-suited to describe the instantaneous motion of moving frames associated with parametric curves. In this thesis, we derive rotors that perform the instantaneous motion of a parametric curve for the translation frame, Bishop frame and Frenet-Serret frame in 3 dimensions. We demonstrate the use of these rotors by iteratively applying them to the corners of a square to construct a square tube fitted around a helix. We also generalize the construction of the rotors associated with moving frames to an arbitrary number of dimensions. We thus further develop the use of PGA in the field of differential geometry.
Distributed Eventual Durability
(University of Waterloo, 2025-08-25) Boulis, Joseph
We present the first design and implementation of a Distributed Eventually Durable transactional system. Durability latency often dominates transaction commit time in distributed databases, limiting performance for latency-sensitive applications. The Eventual Durability (ED) model (VLDB’24) addresses this by decoupling durability from commit, allowing transactions to commit quickly and persist in the background. So far, the ED model has been applied only to centralized databases. Extending ED to distributed, multi-shard settings introduces new challenges, as each shard persists changes independently potentially making partial changes of a transaction durable while other changes may be lost due to failures. This work addresses these challenges effectively by first defining a new correctness criterion, ED Snapshot Isolation (ED-SI), which ensures that transactions only observe the effects of non-failed committed transactions. To enforce ED-SI, we develop a distributed commit protocol that builds on Percolator and adds validation to prevent reads from failed transactions. Our prototype of this system, called ED Percolator, deployed on AWS, supports both fast (speculative) and safe (durable) transactions to allow applications a latency vs. durability trade off while preserving formal correctness. Experiments show that ED fast transactions offer up to 7.6x speed-up over classical Percolator transactions, while ED safe transactions incur similar latency but abort significantly fewer transactions.
Improving Automated Lung Ultrasound Interpretation with Self-Supervised Learning
(University of Waterloo, 2025-08-25) VanBerlo, Blake
Lung ultrasound (LU) is an increasingly important point-of-care examination in acute healthcare settings. In addition to exhibiting comparable accuracy to conventional imaging modalities such as radiography or computed tomography, ultrasound provides safety from radiation and enhanced portability at a reduced cost. Despite its purported benefits, LU has yet to be widely adopted due to a lack of trained experts and education programs. In response, machine learning algorithms have surfaced to alleviate the skills gap by providing automated interpretation. However, the training of machine learning algorithms requires vast amounts of manually annotated examples - that is, images labelled by an expert for whether or not they exhibit particular findings. Given that there are few experts qualified to provide quality annotations, there is a need to explore alternative machine learning techniques for this unique problem. Self-supervised learning (SSL) has emerged as a class of methods to train machine learning algorithms to extract salient features from data without relying on labels. These so-called “pretrained” algorithms constitute better starting points when unlabelled data is abundant but labelled examples are scarce or challenging to acquire, as is the case for LU. Therefore, the purpose of this thesis is to propose and evaluate novel techniques tailored to lung ultrasound that enhance performance on core interpretation tasks, such as detecting lung sliding and pleural effusion. The thesis consists of four studies. The first study explores the efficacy of SSL techniques in brightness mode LU interpretation tasks and proposes a multi-task framework that reuses a single pretrained algorithm to efficiently interpret LU images. The results demonstrate that SSL improves performance on multiple classification tasks. In addition, experiments show that pretrained algorithms amplify generalizability to external healthcare centres. The second study introduces SSL methods to motion mode ultrasound and proposes data augmentation techniques specific to motion mode. We observe that pretrained algorithms achieve the greatest performance on local and external test data for the challenging task of lung sliding classification. The third study proposes a LU-specific technique for SSL that involves sampling pairs of images from the same ultrasound video that are temporally or spatially proximal to each other, based on the intuition that such pairs of images share similar content. The results indicate that, with appropriate parameter assignments, this sampling strategy improves performance of pretrained algorithms on multiple tasks. Lastly, the fourth study proposes ultrasound-specific data augmentation and image preprocessing methods for SSL. The results underscore the value of ultrasound-specific image preprocessing in SSL. A comprehensive evaluation finds that ultrasound-specific data augmentation yields the best performance on a diagnostic task, and that techniques based on cropping attain top performance on object classification tasks. Overall, the findings of this thesis demonstrate that SSL improves the performance of machine learning algorithms for LU, especially on LU images originating from external healthcare centres. Novel SSL methods for LU are established as a key ingredient for producing algorithms that are effective for multiple LU interpretation tasks.
DeTAILS: Deep Thematic Analysis with Iterative LLM Support
(University of Waterloo, 2025-08-25) Sharma, Ansh
Reflexive thematic analysis (TA) yields rich insights but is challenging to scale to large datasets due to the intensive, iterative interpretation it requires. We present DeTAILS: Deep Thematic Analysis with Iterative LLM Support, a researcher-centered toolkit that integrates large language model (LLM) assistance into each phase of Braun \& Clarke’s six-phase reflexive TA process through iterative human-in-the-loop workflows. DeTAILS introduces key features such as “memory snapshots” to incorporate the analyst’s insights, “redo-with-feedback” loops for iterative refinement of LLM suggestions, and editable LLM-generated codes and themes, enabling analysts to accelerate coding and theme development while preserving researcher control and interpretive depth. In a user study with 18 qualitative researchers (novice to expert) analyzing a large, heterogeneous dataset, DeTAILS demonstrated high usability. The study also showed that chaining LLM assistance across analytic phases enabled scalable yet robust qualitative analysis. This work advances Human-LLM collaboration in qualitative research by demonstrating how LLMs can augment reflexive thematic analysis without compromising researcher agency or trust.
Unifying Foundation Models with Decision Making: A Path Towards Autonomous Self-Improving Agents beyond Rewards and Human Supervision
(University of Waterloo, 2025-08-25) Chandra, Abhranil
Recent progress in AI has primarily been driven along two axes - scaling large generative Foundation Models (FMs) trained on static internet scale data and developing better sequential decision making (DM) and reinforcement learning (RL) algorithms that allow experiential learning. Both paradigms alone have their own flaws, but together provide a scalable recipe towards general intelligence. As AlphaGo made its famous "Move 37", it reaffirmed Reinforcement Learning's(RL) efficacy as a paradigm to bootstrap intelligence via interactive learning from scratch, optimizing for goal-directed behavior, self-improvement, and emergence of novel superhuman abilities. However, RL is hard to scale beyond narrow tasks or simulated environments given the scarcity of real-world decision-centric data, sparsity of feedback, hard-to-design rewards, and problems of scaling to larger models. On the contrary, recent generative FMs pretrained on static internet-scale text and image data excel at broad high-level knowledge acquisition but fail to acquire robust internal World Models, precluding their notion of agency, ability to plan and reason well, and extrapolation beyond training data. In the thesis we focused on combining these complementary paradigms - to both build and use FMs for DM and develop DM and RL tools for improving FMs - using DM/RL as a paradigm to finetune and optimize general-purpose pretrained models to elicit better decisions beyond training data, rather than viewing it as a paradigm to bootstrap intelligence from scratch. Such broadly capable systems can then be used to empower agents that can perceive, reason, and act robustly in both physical settings to complete embodied tasks and in virtual settings to help as autonomous task completion and knowledge-discovery agents. First, we introduce VideoAgent, a jointly trained goal-conditioned video generation policy and self-improving simulator for embodied planning tasks. VideoAgent learns to refine its own generated plans based on a novel self-conditioning consistency objective and also using feedback from pretrained vision-language models (VLMs), without requiring ground-truth action labels or any explicit rewards. The model further leverages search to enable iterative improvement of the video plans using inference-time compute, leading to more grounded and physically plausible plans in robotic manipulation tasks. Second, we develop a framework, Reg-ReBoot, to investigate efficient and scalable methods to improve base non-reasoning LLMs into better reasoners without using explicit verified data or rule-based verifiers. We analyze this counterintuitive idea: fine-tuning language models using unverified and even incorrect reasoning traces leads to reasoning improvement. We show that large language models (LLMs), due to their inductive biases, can learn useful reasoning heuristics by averaging over noisy chain-of-thought (CoT) data. Our results on mathematical reasoning benchmarks reveal that noisy synthetic data can be an efficient way to bootstrap performance and decrease reliance on hard to curate verified solutions. From these insights we propose a two stage mid-training pipeline - lowering the barrier to scalable reasoning improvement. Finally, we address the evaluation bottleneck in generative modeling by proposing ReFeR, a multi-agent, tuning-free evaluation framework. ReFeR uses a hierarchy of pretrained LLMs and VLMs to provide automatic, scalable, and high-quality feedback for both textual and visual generations. This framework not only rivals human-level evaluation accuracy but also produces structured feedback, enabling downstream distillation and fine-tuning, and proving effective even in complex reasoning tasks. Together, through these works, I try to contribute towards the goal of building autonomous self-improving agents: one where systems powered by foundation models leverage test-time compute, generative simulations and world models, and diverse learning signals beyond explicit rewards and human feedback to drive interactive learning and decision making.

Browse

Recent Submissions