21090 research outputs found
Sort by
JPEG-Inspired Encoding for Deep Learning
No full textJPEG is the dominant standard for storing and transmitting digital images, while Deep Neural Networks (DNNs) have become the preeminent method for automated image understanding. This dissertation investigates how these two ubiquitous technologies can be synergistically integrated to enhance the performance of DNNs. JPEG was originally engineered for the Human Visual System (HVS), and its default parameters are not optimized for DNNs, which process visual information differently. This suboptimality, stemming from JPEG’s default implementation, is not a fundamental limitation but rather an opportunity to adapt its core components—especially the non-linear quantization stage—for DNNs. This research addresses this suboptimality by first optimizing the trade-off between compression rate and classification accuracy, and second, by introducing a learnable, end-to-end differentiable JPEG layer whose quantization parameters are jointly trained with the underlying DNN. This dissertation demonstrates that this principle of a learnable, JPEG-inspired transformation extends beyond compression, offering a novel way to address challenges in related domains such as knowledge distillation (KD), where large 'teacher' models often overfit the training set. This overfitting causes them to generate overconfident, near one-hot probability vectors that serve as poor supervisory signals for the student model, suggesting the need for novel approaches to information transfer. This dissertation addresses these issues by systematically revisiting the relationship between JPEG encoding and deep learning. It charts a logical progression from adapting JPEG externally for DNNs, to integrating it internally as a learnable network component, and finally to repurposing its core principles to amplify knowledge transfer.
This progressive framework is methodically developed and empirically substantiated through three interconnected contributions:
-Optimizing Compression for DNNs.
To improve the interaction between standard JPEG and pre-trained DNNs, this work first reframes compression from a human-centric "rate-distortion" problem to a DNN-centric "rate-accuracy" one. This is achieved by introducing the Sensitivity Weighted Error (SWE), a novel distortion measure derived from a DNN’s loss sensitivity to frequency-domain perturbations, where higher sensitivity in a frequency band indicates its greater importance for the DNN’s decision-making. The SWE guides the OptS algorithm to generate model-specific JPEG quantization tables. This approach produces fully compliant JPEGs optimized for DNN consumption, demonstrably improving the rate-accuracy trade-off by increasing accuracy up to 2.12% at the same rate, or enabling rate reductions up to 67.84% with no loss of model accuracy.
-Integrating a Differentiable JPEG Layer into the DNN Architecture.
Building on this, the next contribution integrates the codec into the network architecture itself via the JPEG-Inspired Deep Learning (JPEG-DL) framework, which introduces a novel, end-to-end differentiable JPEG layer. By replacing JPEG's standard hard quantization with a differentiable alternative, this layer's parameters are jointly optimized with the network's weights. This transforms the JPEG pipeline from a static pre-processor into a dynamic, learnable component, significantly improving model accuracy—by an average of 7% on fine-grained classification tasks with only 128 additional trainable parameters—and enhancing robustness against adversarial attacks.
- Amplifying Knowledge Transfer via JPEG-Inspired Perturbation.
Finally, the differentiable layer is repurposed to address the "overconfident teacher" problem in KD by perturbing teacher inputs to force softer, more informative predictions. Crucially, this method requires no retraining or modification of the fixed teacher model, ensuring its practical utility with proprietary or deployed networks. Our investigation begins with Coded Knowledge Distillation (CKD), a practical heuristic that uses adaptive JPEG compression to perturb teacher inputs and soften their overconfident predictions. While effective, this approach prompted a search for a more principled theoretical foundation. This led to Generalized Coded Knowledge Distillation (GCKD), a framework that establishes the maximization of the teacher's Conditional Mutual Information (CMI) as the core objective. However, directly optimizing for CMI on a per-input basis is computationally prohibitive. This efficiency challenge is resolved in the culminating synthesis, Differentiable JPEG-based Input Perturbation (DJIP). DJIP operationalizes the GCKD theory by deploying the trainable differentiable JPEG layer as a fast, learnable, and amortized operator. Instead of performing a slow, per-input optimization search, the layer is trained once to automatically generate CMI-maximizing perturbations, making the process highly efficient. This approach demonstrably generates richer supervisory signals, boosting student model accuracy by up to 4.11%.
In conclusion, this dissertation demonstrates that the relationship between JPEG and DNNs can be systematically revisited to create a powerful synergy. By progressing from adaptation to integration and synthesis, this work transforms the suboptimal default interaction of JPEG and DNNs into a versatile architectural tool. The research delivers a suite of methods that not only improve the performance of DNNs on compressed images but also offer a theoretically-grounded solution to a key challenge in knowledge distillation. By demonstrating that legacy codecs can be repurposed to enhance model accuracy, efficiency, and knowledge transfer, this work thus reframes the role of classical codecs, proposing JPEG-inspired encoding as a principled foundation for the integration of classical compression and deep learning
Monastic Diets and Aquatic Species: Examining Potential Fish Consumption at the Ghazali Monastery, Sudan Through Stable Isotope Analysis of Sulphur
No full textStable isotope analysis can be applied in bioarchaeological contexts as a tool to assess paleodiet as this technique relies on naturally occurring differences in isotopic values in different food sources and environments. Previous research has been conducted to assess possible dietary composition of the monastic inhabitants of at the medieval Makurian site of Ghazali (ca. 680-1275 CE), Nubia using stable isotope analyses of carbon (δ13C) and nitrogen (δ15N) on bone collagen of its monastic inhabitants. This showed a varied diet evidently comprising of both terrestrial plants and animals. Additionally, δ15N values observed in five individuals suggest potential aquatic species consumption in conjunction with terrestrial animal protein. However, no remains of aquatic species were identified during excavations at Ghazali, and little aquatic species were identified at other Makurian sites. This lack of evident aquatic species (e.g. fish) consumption at Ghazali brings forth numerous questions surrounding dietary practices both at Ghazali and within the broader region of similar Makurian monasteries.
This research utilized stable isotope analysis of sulphur (δ34S) on human bone collagen in conjunction with previously presented δ15N values, in tandem with existing textual and bioarchaeological evidence from Egypt and Byzantium, to determine the presence or absence of fish in the diet of the monastic inhabitants at Ghazali. The sample consisted of 20 individuals from Cemetery 2, where 18 of these individuals were male monks. Analysis of δ34S, when coupled with previous δ15N values, revealed that four of these individuals showed evidence of possible fish consumption alongside terrestrial animal protein consumption
Rewriting the City: Disruption as Cultural and Spatial Practice in Amman, Jordan
No full text“Rewriting the City” examines how everyday informal spatial practices in Amman—such as street vending, ad-hoc construction, and mobile economies—are framed as disruptions by state-led planning and municipal authorities, while functioning as strategies of livelihood for those who depend on them. Operating outside formal architectural authorship, these practices remain integral to how the city is inhabited, serviced, and sustained.
The thesis investigates how such practices come to be labelled as “disruptive” within planning and regulatory discourse, and how this designation frames them as problems of order rather than as intentional spatial practices. It argues that disruption is not an inherent quality of these acts, but a relational one: they are deemed disruptive only within systems that prioritize legibility, permanence, and control. Therefore, what is disrupted is not urban order, but dominant frameworks that define architecture, legality, and spatial authorship.
Focusing on markets, sidewalks, streets, and mobile economic infrastructures across East and West Amman, the research analyses how informal practices operate within, alongside, and against municipal planning and governance. The East-West divide is approached as a spatial condition produced by uneven infrastructure investment, zoning practices, and histories of migration and displacement, revealing how regulation itself produces uneven access, visibility, and legitimacy.
The thesis reframes disruption as persistence: the repeated, adaptive occupation of space that sustains livelihood under constraint. These practices actively shape public space, organize social relations, and maintain systems of interdependence through deliberate responses to uneven development, economic precarity, and regulatory constraint.
Methodologically, the thesis employs ethnographic research, drawing on observation, storytelling, and spatial documentation to examine how space is transformed through temporality, mobility, material decisions, bodily labour, and repeated use.
By reading these practices as architectural, the thesis expands the discipline’s scope and responsibility. It challenges the association of architecture with permanence, capital, and professional authorship, and instead positions architecture as a contingent and relational practice embedded in agency, labour, and lived experience
Reimagining High School: A Guide to Renewing Post-War Secondary Schools in Waterloo Region
No full textOntario’s public education system is in a state of crisis; this is evident in the architecture of schools. Many of Ontario’s secondary schools were built during the post-war era (1955-1975), and education continues to operate within buildings that reflect the values and priorities of that time. Education during this period was highly institutionalized, and buildings were designed with the intention of enforcing rigid standards of learning and behaviour. Classrooms prioritized control and uniformity to administer required tests and performance assessments effectively. While funding has been directed towards additions and band-aid solutions, these schools have become crowded over time, consequently relying on portable classrooms. Despite shifts in educational values and pedagogy, the architecture of secondary schools has remained largely unchanged for the past fifty years, failing to meet the priorities and evolving pedagogy of today.
Since the post-war era, educators and policymakers have acknowledged that each student and school community has unique needs. While recent pedagogical advancements have been successfully integrated into well-funded, progressive, and newly constructed schools, there is still a gap in understanding how existing infrastructure, particularly post-war schools, can be upgraded to support these educational principles. The thesis research proposal seeks to develop a framework for reimagining secondary education by renewing an existing post-war secondary school in the Waterloo Region, Ontario. The proposal aims to support modern educational practices, embracing 21st-century learning, diverse learning styles, and community-based education. To guide this research, three interconnected scales are examined: the classroom, the school building, and the surrounding urban context.
The research objectives are to assess the current state and future vision of public education in Ontario, understand the evolving needs of 21st-century learners, and explore how schools can support both student development and community well-being. The research culminates in a set of design guidelines that outline strategies for renewing existing school buildings into inclusive, adaptable, and community-oriented environments that reflect contemporary educational values. The primary guiding questions are: How can post-war secondary schools, originally designed under outdated educational philosophies, be adapted for 21st-century learning? In what ways can spatial design foster a sense of belonging, engagement, and well-being among students? How can the schools establish a closer connection with their surrounding communities
Novel Class Discovery for 3D Point Cloud Semantic Segmentation in Large-Scale Environments
No full textModern urban environments undergo continuous transformation as emerging infrastructure appears worldwide. Traditional semantic segmentation methods for 3D point clouds operate on fixed taxonomies, producing static representations that cannot adapt to novel categories. This dissertation addresses novel class discovery (NCD) in large-scale 3D point cloud segmentation through geometry-aware mechanisms, adaptive multi-source fusion, and hybrid supervision frameworks.
The first study establishes geometric foundations through voxel-geometry integration with region-centric organization, termed CHNCD. The framework couples voxel representations with original spatial coordinates via index mapping, identifies semantically informative points within clusters, accelerates neighbor retrieval through proximity hash mapping, and consolidates localized features with global context via spatial attention. Experiments on S3DIS, Toronto-3D, SemanticSTF, and SemanticPOSS demonstrate consistent improvements over discovery baselines.
The second study deepens representation through adaptive geometric sequence modeling, dynamic Gaussian embeddings, and gated multi-source fusion, termed AGDNet. Adaptive geometric sequence modeling employs learnable dimension weighting and dynamic grouping adjusted to local point density. Dynamic Gaussian embeddings represent point clouds as 3D Gaussians and compute Mahalanobis distances to generate multi-scale spatial embeddings. Gated multi-source fusion intelligently weights features through context-aware mechanisms. Three knowledge-transfer objectives operate at category, instance, and distribution levels to bridge semantic gaps. Evaluation on Toronto-3D, SemanticSTF, and SemanticPOSS demonstrates substantial improvements.
The third study integrates discovery with operational land cover mapping, termed 3DLCDM. The framework processes features through a supervised head for established categories and a dual unsupervised head comprising a primary branch with fixed prototypes and an over-segmentation branch with progressive scheduling. Temporal Sinkhorn-Knopp normalization with adaptive temperature scheduling stabilizes pseudo-labels, while dynamic weighting combines per-batch and global frequency statistics to address class imbalance. Evaluation on DALES and H3D datasets demonstrates substantial improvements for continuous land cover discovery mapping.
Taken together, the three studies advance a progressive research agenda unifying discovery and 3D segmentation for large-scale point cloud scenes. The dissertation demonstrates consistent gains across six benchmark datasets, exhibits generalization across sensors and acquisition geometries, and provides a principled route to maintain updated urban maps as new structures emerge
Considerations for Commons Governance in Chilika Lagoon: New-Commonisation through Codification
No full textThis thesis examines how communities can re-establish governing authority over shared environmental resources (commons) after periods of dispossession (decommonisation), a process described as ‘new-commonisation’. Focusing on Chilika Lagoon, India, it explores how small-scale fishery communities might regain autonomy following decades of externally-driven decommonisation, caused by privatization, encroachment, elite capture, and fragmented state interventions. The central argument is that legally-grounded recognition of commons is helpful for re-gaining rights and essential for protecting communities from renewed external threat. Drawing on process-tracing analysis of three cases; Shimshal Valley in Pakistan, forest governance under India’s Forest Rights Act (2006), and Locally Managed Marine Areas (LMMAs) in Papua New Guinea; the study identifies how community mobilisation and legal codification interact to regain and stabilize self-managed commons. Though the findings are hypothesis generating rather than hypothesis testing, they suggest that enduring governance outcomes emerge when communities achieve de jure recognition of de facto rights, and that their success depends on contextually-dependent enabling conditions, such as equitable enforcement, multi-level support and the mechanism for legal rights. As no two commons are identical, there is no single path to codification; legal arrangements must respond to the specific socio-political and ecological context of each community. This research contributes to commons theory by framing codified legal backing as a critical, yet under-developed, dimension of enduring commons governance, in the face of persistent external pressures
Model-Based Optimization of pH and Temperature in Chinese Hamster Ovary Cell Culture
No full textMonoclonal antibodies (mAbs) are widely produced in mammalian cell cultures, with Chinese Hamster Ovary (CHO) cells being the predominant host cell used in the pharmaceutical industry. The growing global demand for mAbs has driven significant advances in biomanufacturing and motivated the pharmaceutical sector to develop strategies that enhance productivity. Among the key factors influencing mAb yield are the operating conditions of CHO cell cultures, such as pH and temperature. Optimizing these parameters is therefore essential for improving process performance and product quality.
Model-based optimization offers a powerful and systematic approach for improving complex bioprocesses, including mAb production. Combining mechanistic understanding with mathematical modeling, it enables quantitative prediction of process behavior and identification of optimal operating strategies without excessive experimentation. In essence, model-based optimization relies on two critical components: (1) a dynamic model capable of accurately describing process behavior, and (2) an optimization algorithm that determines the best operating conditions based on model predictions. The effectiveness of a Model-based optimization depends on both components working reliably to ensure convergence toward realistic and true optima.
The repetitive nature of batch processes makes them particularly suitable for batch-to-batch optimization, where information from previous runs is used to improve future ones. In this iterative framework, process measurements from a completed batch are used to update the model and compute the optimal input profile for the next experiment. However, optimization of mammalian cell cultures is challenging because of the strong nonlinearities and interactions among growth, metabolism, and product formation under varying environmental conditions. These complexities often lead to model–plant mismatches, so parameters estimated through model identification may not accurately reproduce the true gradients of the cost function or constraints, which are quantities that are essential for optimization.
To address this, a modified batch-to-batch optimization, so-called the simultaneous identification and optimization method, is employed. This approach forces the model-predicted gradients to match experimentally measured gradients by adjusting model parameters, while an output correction term ensures that previously achieved fitting accuracy is retained. Consequently, the resulting parameter set satisfies both identification and optimization objectives even when structural model errors are present. Despite its potential, several challenges must be overcome before applying this framework to complex biological systems. Previous studies have computed and corrected gradients only at the end of each batch; however, incorporating transient, within-batch measurements could provide richer information and improve the characterization of model discrepancies. Additionally, integrating optimal experimental design can enhance parameter identifiability and accelerate convergence, and the framework can be further extended to continuous operation modes. Most importantly, the methodology has not yet been thoroughly tested in a real experimental system to demonstrate its performance and robustness.
A reliable mechanistic model capable of describing CHO cell metabolism under varying process conditions is also essential but remains insufficiently explored. Two major modeling paradigms exist for bioprocesses: kinetic models and dynamic flux balance analysis (dFBA). Kinetic models employ ordinary differential equations to relate measurable process variables, such as viable cell density, substrate and by-product concentrations, and product titer, to underlying rates of growth, uptake, and synthesis. In contrast, dFBA models optimize a biological objective (e.g., growth rate) with respect to intracellular fluxes subject to stoichiometric and steady-state constraints. Compared to purely kinetic models, dFBA frameworks can offer deeper physiological insight and require fewer parameters, making them particularly attractive for model-based optimization and control.
Building upon these concepts, this thesis first presents a novel dFBA model integrated with kinetic constraints to predict the dynamic metabolism of CHO cells under varying pH and temperature conditions in fed-batch cultures. The model captures the main metabolic behaviors across different operating conditions. In the subsequent chapters, the batch-to-batch optimization framework is extended and modified for both batch and continuous bioprocesses, incorporating gradient correction and optimal experiment design to ensure robustness and faster convergence. Finally, the developed methodology is implemented and experimentally validated using an AMBR-15 mini-bioreactor system, where it is applied to determine the optimal pH profile that maximizes monoclonal antibody production in CHO cultures
Designing Public Health Surveillance for Urban Air Quality in LMICs: Community Insights, Technology Acceptance, and System Design for Low-Resource, High Vulnerability Settings
No full textClimate change is tightening exposure windows and widening inequalities in urban air quality, especially across low- and middle-income countries. Many cities lack dense regulatory networks, timely analytics, and trusted communication pathways, which means signals arrive after decisions are due. Grounded in Ulaanbaatar, Mongolia, this thesis begins by asking how people make sense of pollution in their daily lives, what actions are realistically available, and which institutions are expected to respond. These lived accounts specify what usable guidance must deliver in contexts where resources are limited and risks are uneven. Guidance must be fast, intelligible, transparent about uncertainty, and aligned with social roles and constraints that vary.
A second qualitative strand examines technology acceptance of digital monitoring and early warnings. It identifies what confers legitimacy, including credible data provenance, visible accountability, and delivery pathways that match capabilities such as low connectivity, limited time, and competing obligations. Together, these qualitative insights establish system requirements and the conditions under which guidance is likely to be acted upon.
Based on these insights and in partnership with UNICEF Mongolia, the thesis designs, develops, and evaluates a real-time air quality pipeline for Ulaanbaatar. Low-cost sensors feed an automated device to database workflow that stabilizes sparse and noisy inputs. A sequence modeling approach produces continuous predictions with calibrated error suitable for communication and decision support under intermittent power and limited connectivity. Evaluations suggests the system performs reliably under these constraints and can be adopted within existing civic workflows.
The integrated contribution is a pathway from qualitative insights to deployable infrastructure that supports proportional protection. The thesis advances empirical understanding of disproportionate risks in an LMIC city, delivers a validated and operational monitoring and prediction pipeline build from locally derived requirements, and offers policy and design guidance that ties technical accuracy to local relevance and shared accountability so that evidence arrives in time to reduce harm
Categories as a Foundation for both Learning and Reasoning
No full textThis thesis explores two distinct research topics, both applying category theory to machine learning.
The first topic discusses Vector Symbolic Architectures (VSAs). I present the first attempt at formalising VSAs with category theory. VSAs are built to perform symbolic reasoning in high-dimensional vector spaces. I present a brief literature survey demonstrating that the topic is currently completely unexplored. I discuss some desiderata for VSA models, then describe an initial formalisation that covers two of the three desiderata. My formalisation focuses on two of the three primary components of a VSA: binding and bundling, and presents a proof of why element-wise operations constitute the ideal means of performing binding and bundling. The work extends beyond vectors, to any co-presheaves with the desired properties. For example, GHRR representations are captured by this generalisation.
The second line of work discusses, and expands upon, recent work by Milewski in the construction of "pre-lenses." This work is motivated by pre-established formalisations of supervised machine learning. From the perspective of category theory, pre-lenses are interesting because they unify the category Para, or Learn, with its dual co-Para, or co-Learn. From a computer science perspective, pre-lenses are interesting because they enable programmers to build neural networks with vanilla function composition, and they unify various network features by leveraging the fact that they are profunctors. I replicate Milewski's code, extend it to the non-synthetic data, MNIST, implement re-parameterisations, and describe generative models as dual to discriminative models by way of pre-lenses. This work involved creating a simple dataloader to read in external files, randomising the order that inputs are presented during learning, and fixing some bugs that didn't manifest when training occurred on the very small dataset used by Milewski
Leveraging long term water quality monitoring data to elucidate drivers and controls on N and P loading in the Lake Winnipeg Basin
No full textSince the late 1920s, dams have been a common tool used in the Lake Winnipeg basin to control flooding during spring snowmelt and to supply freshwater during the summer dry season. As a result, there are now over 140 reservoirs larger than 10 ha in the watershed. Many retain a substantial fraction of the inflowing riverine phosphorus (P) load, and some also remove inflowing nitrogen (N). Yet we lack a quantitative understanding of how reservoir nutrient removal reshapes seasonal nutrient delivery and N:P stoichiometry at the watershed scale. Removal efficiency variation, both between reservoirs and interannually, and its driving factors also remain poorly constrained. Chapter 1 gives an overview of the eutrophication challenges facing Lake Winnipeg, puts the challenges into context based on the Lake Winnipeg Basin’s land use, water management and climate, and summarizes how we expect reservoirs are influencing TN and TP loads across the watershed.
In Chapter 2 we leverage 13 case-study reservoirs where flow and concentrations of total phosphorus (TP) and total nitrogen (TN), have been monitored near the inlet(s) and outlet over multiple years. Using the Weighted Regression on Time Season and Discharge model with Kalman filtering (WRTDS-K) we calculate P and N loads, estimate TP retention and TN removal efficiencies, and examine daily model coefficients to quantify how reservoirs modify TP and TN concentration-discharge (C-Q) relationships throughout the year. We further investigate flow volume, intensity and timing as potential drivers of interannual variation in TP retention and TN elimination efficiencies. On average, reservoirs retain 43.9% ±35.2% of inflowing TP and eliminate 6.98%± 30.9% of inflowing TN, systematically increasing the N:P ratio of downstream nutrient loads. They also consistently suppress the C-Q slopes for both TN and TP, with most pronounced impacts occurring during the spring freshet. Multiple linear regression analysis demonstrates that metrics of flow timing can explain 10-50% of inter-annual variability in retention efficiency. Collectively, these findings demonstrate that reservoirs in the Lake Winnipeg basin systematically reshape the magnitude, timing and stoichiometry of nutrient exports, with important implications for nutrient-management strategies under a changing climate.
In Chapter 2, I showed that reservoirs have a large impact on the in-stream concentrations and loads of total phosphorus (TP) and total nitrogen (TN) across the Lake Winnipeg Basin (LWB). However, only a small fraction of the 140+ reservoirs in the basin have sufficient paired inflow and outflow monitoring data to accurately determine their contributions to TP and TN retention or enrichment. Based on the thirteen case study sites considered in Chapter 2, Chapter 3 uses Generalized Additive Models (GAMs) to evaluate the relative importance of a broad range of predictor variables representing reservoir and environmental factors that potentially influence the efficiency of removal of TP and TN from reaching Lake Winnipeg. The GAM features are evaluated by stepwise adding factors to the model and determining the associated improvements of performance metrics. The best model is then further assessed by comparing the residuals and adjusted R2 values generated from a leave-one-out cross-validation. The results emphasize the added value of incorporating information on water chemistry and C-Q relationships when assessing reservoir influence in water quality models, improving the accuracy of predictions of the impact of reservoirs on TP and TN concentrations and loads, especially when considering their interannual variability. The insights from Chapter 3 also contribute to a broader understanding of how reservoirs in general control riverine TP and TN flows