21793 research outputs found
Sort by
Northern Lights and Silicon Dreams: AI Governance in Canada (2011-2022)
No full textBased on a three-year investigation into national and provincial AI governance, the report’s major findings include:
Canadian AI governance focuses on economic and industrial policy. National symbolic investment in AI – the promotion of AI’s Canadian-ness – impedes critical discussion about the technology and its risks.
AI governance is uncoordinated and lacks clear mandates for consultation and effective mechanisms of feedback, which impedes good governance and public participation.
AI policy is marked by notable silences on key issues, including Indigenous rights and data sovereignty, the creative and cultural sectors, and the environmental impact of AI.
The Government of Canada is a key site of AI development and deployment that remains understudied in current legislation.
We reached these findings based on investigations into a range of topics, including the development of the Canadian government’s algorithmic impact assessment, the way ethical concerns over AI are circumnavigated during the procurement process, flawed public consultations on the use of facial recognition technologies, the history of the Canadian Institute for Advanced Research’s involvement in AI research and the administration of national AI policy, and the controversial use of racially biased AI technology by Immigration, Refugees and Citizenship Canada.
The views of each chapter are of the authors themselves. Collectively, the report foregrounds the need to improve AI consultation and public participation in AI governance. The chapters also throw into question the assumptions and efforts that led to the development of the Artificial Intelligence and Data Act and its inclusion in the larger implementation of Bill C-27. The passage of this act raises concerns not just about AI but about digital policy development and, indeed, the fragile state of democratic governance, accountability, and oversight in Canada.
Shaping AI is part of a multinational and multidisciplinary social research project that examines the global trajectories of public discourse on AI in four countries (Germany, UK, Canada, and France) over a ten-year period, 2012-2021. Funded by the European Open Research Area initiative for a period of three years (February 2021 – February 2024), Shaping AI brings together leading research teams from each of the four countries under scrutiny.
Research from this report was coordinated by Dr. Sophie Toupin and Dr. Fenwick McKelvey. This report compiles research conducted as part of the Shaping AI Canadian policy research activities that ran from 2021 to 2023.
Thanks to Blair Attard-Frost and Aaron Tucker for providing peer review of the report.
This edited collection draws on research supported by the Social Sciences and Humanities Research Council
this cloud, this crust, this doubt, this dust
Get PDFthis cloud, this crust, this doubt, this dust is an exhibition of work about my obsession with Virginia Woolf, her novels, and her home. It is also accompanied by a written text exploring skies, skins, ambiguity, and remainders
Automated Progress Monitoring and Reporting for Construction Projects
Get PDFIn complex and dynamic construction sites, efficient progress monitoring and reporting play an important role in minimizing schedule delays and cost overruns. Such reporting requires detailed and accurate records from job sites to help project managers in comparing project’s current state to its as-planned state. Manual traditional progress reporting is time-consuming, costly, labour-intensive, and error-prone. In recent years, advancements in technologies and methods have been introduced in an effort to overcome the challenges of manual methods and to automate the processes of progress monitoring and reporting. These introduced levels of automation still lack capabilities to provide complete and accurate information about the project’s current status and available resources on job sites. To address these challenges, this thesis introduces a novel framework for automated progress reporting in construction. This framework provides detailed information for each tracked building element, enabling the identification of its current status and the generation of timely progress reports. The developments integrated into the framework focus on challenges associated with congested mechanical components in indoor environments. Monitoring these components is crucial because their complex and time-consuming installation procedures can lead to project delays.
The developed framework consists of three main modules: (i) Object Recognition (ii) Object Localization, (iii) Integrated Object Recognition and Localization. In the “Object Recognition” module, two deep learning algorithms, YOLACT++ and Mask R-CNN, were utilized in processing digital images captured at construction sites for the automated recognition of tracked building elements. YOLACT++ proved superior to Mask R-CNN and was accordingly utilized in the developed framework. In the “Object Localization” module, a Real-time Locating System (RTLS) is utilized to identify the location of each recognized element along with its ID. The Ultra-wideband (UWB) system was selected as an RTLS, and different laboratory and field experiments were conducted to validate the UWB system’s localization performance. Finally, in the “Integrated Object Recognition and Localization” module, a user-friendly application was developed to integrate the outputs from the YOLACT++ model and the UWB system and automatically generate status reports of tracked elements. These reports include visual and location information, along with the unique ID of each element.
The framework was tested and validated using 3,632 images. The results demonstrate good performance and effectiveness of the developed framework under challenging conditions; yielding recognition accuracy of close to 85% in precision and recall for HVAC duct and slightly less than that for pipes. Similar performance was achieved in localization, yielding errors ranging from 0.03 to 1.22 meters in two-dimensional (2D) coordinates and from 0.15 to 1.6 meters in three-dimensional (3D) coordinates in the field test. The developed framework can be easily extended to other building elements, and the excel format of its output can facilitate linkage with Building Information Modeling (BIM) systems
Using augmented reality to improve pre-surgical decision making among breast cancer patients
Get PDFMost breast cancer patients will undergo surgery as part of their treatment plan. To improve their chances of survival, they typically need to decide on their treatment plan within eight weeks of diagnosis. Patients often review post-operative images of others to gauge potential outcomes, but these images provide only limited insight into their own possible results resulting in many revision surgeries and patients who are dissatisfied with their surgical outcomes.
In this dissertation, we explore the use of augmented reality (AR) as a decision-support tool to help patients visualize different surgical procedures on their own bodies. To that end, we developed Breamy, an AR app that uses both marker-based and markerless AR visualizations. Breamy uses photogrammetry to create a patient-specific 3D model. This model, along with various treatment options, is projected directly onto the patient’s body to help visualize different surgical options.
We surveyed 165 women on their views about the concept of Breamy and found positive results in terms of the need for such an application. We also ran a preliminary study with six participants to evaluate the usability of Breamy and its potential as a decision-aid tool. The findings of these studies suggest that AR can be an effective decision-support tool, helping to better align patient expectations with likely outcomes. For example, in our preliminary study, 90\% of participants believed that an AR application with personalized surgical information could improve patient comprehension and decision-making. Similarly, in our second study, five out of six participants reported that AR visualization enhanced their understanding of surgery's potential effects on their bodies and boosted their confidence in the decision-making process. The results of our studies underscore the potential of Breamy to transform the surgical decision-making process, ultimately leading to greater patient satisfaction and improved surgical outcomes
Development of a Condition Assessment Rating System and Prediction Model for Railway Tracks
Get PDFCanada has an extensive rail network spanning 45,000 kilometres. The railway system plays a
crucial role in serving almost every sector of the Canadian economy. Primarily, it transports freight
to and from the U.S. and global markets through coastal ports. However, failures in the railway
infrastructure can have severe safety and financial consequences. In 2023, 43.13% of main-track
derailments were attributed to track defects, according to the Transportation Safety Board of
Canada. These defects, including issues with track geometry and component failures, underline
the need for better track condition monitoring and maintenance to prevent derailments. This
research aims to address this need by developing a comprehensive rating system for evaluating the
condition of ties and rail fastening components and machine learning models to predict future track
conditions. While traditional condition assessment ratings have relied on subjective evaluations
and considered components separately, this study proposes a Tie and Rail Fastening system that
evaluates the condition of ties, tie plates, and spikes. Domain expertise was incorporated through
the Analytic Hierarchy Process (AHP) to prioritize the importance of various defects. The resulting
weighting system provides a more detailed and integrated approach compared to existing rating
methods, which primarily focus on crack size. Machine learning models, including Random
Forest, XGBoost, and Cat Boost, were employed to predict future conditions, such as defect tags,
amplitude, and length. These models achieved a 95% accuracy for detecting defect tags and a 75%
accuracy when predicting defect tags based on predicted amplitude. On the one hand, the proposed
tie and rail fastening rating system can improve the prioritization of future rail maintenance works.
On the other hand, the proposed machine learning models can improve the planning of future
maintenance by offering better tools for monitoring and predicting track conditions
Advancements in model combination and uncertainty quantification with applications in actuarial science
Get PDFIn this thesis, we focus on model combination, incorporating elements of uncertainty quantification to address different actuarial science issues. We first tackle the issue of overconfidence from a single model combination approach, highlighting how different combination assumptions can lead to different conclusions about the predicted variable. This is illustrated with an extreme precipitation example for the regions of Montreal and Quebec. We then focus on Bayesian model averaging (BMA), a very popular model combination technique relying on Bayes' theorem to attribute weights to models based on the likelihood that the observed data comes from the models considered. We propose a correction to the classical expectation-maximisation algorithm to account for data uncertainty, where we assume that the observed data is in fact not the only possible observable data. We then generalise our method to include Dirichlet regression, allowing for combination weights to vary depending on risk characteristics. These BMA approaches are applied to a simulation study as well as a simulated actuarial database and are shown to be very promising, as they allow for a more formal model combination framework for combining actuarial reserving methods in a smooth way based on predictive variables. Next, we adapt Bayesian model averaging using Generalised Likelihood Uncertainty Estimation to extreme value mixture models, and show that this modification allows for identifiying the "best" extreme value threshold, although a combination of models will outperform the single best mixture model. This is illustrated using the Danish reinsurance dataset. Finally, we show that the generalised BMA algorithm can be used to identify flexible extreme value thresholds depending on predictive variables. We use this generalised mixture model combination on a recent dataset from a Canadian automobile insurer
Optimized Multi-Agent Deep Reinforcement Learning for Target Search and Localization
Get PDFIn a world that increasingly relies on autonomous systems, swarm robotics hold the promise of revolutionizing how complex tasks, such as target search and localization, are approached. The ability of multiple autonomous agents, such as robots and UAVs, to work together, exchange information, and adapt to dynamic environments is critical for target localization applications ranging from search and rescue missions to environmental monitoring. However, efficiently coordinating a swarm of robots to search for targets in uncertain and complex environments poses significant challenges. Most existing solutions for target search and localization still possess challenges and limitations in terms of adaptability to different environments and scalability. These challenges become even more intricate when the target may not exist (i.e. false alarms) or is unreachable.
In this thesis, the main motivation is to leverage AI, specifically Multi-Agent Deep Reinforcement Learning (MDRL), to address the target search and localization problem. The aim is to develop MDRL solutions where the agents intelligently and autonomously learn to tackle the problem and its different complexities, with minimum human intervention. The capacity of MDRL in producing agents capable of learning from their experiences in the environment proves efficient in handling complex and dynamic scenarios, such as coordinating with other agents, translating data readings into actions that lead to the target, and navigating obstacles in the environment. This research is motivated by four main needs: (1) Adaptable solutions for collaborative target search and localization for varying environment complexities; (2) autonomous and intelligent sensing agents with decision-making that addresses scenarios like false alarms and target unreachability; (3) scalable and efficient AI-based learning process for the sensing agents, and (4) mechanisms for knowledge exchange across different users and parties from different domains for better accessibility to AI solutions.
The aforementioned needs are addressed in this thesis by: (1) Developing novel MDRL algorithms for collaborative target search in both simple and cluttered environments by modeling the problem as a Markov Decision Process (MDP), (2) designing novel methods based on ideas from MDRL, Imitation Learning (IL), and reward shaping for enhanced and quick learning performance, (3) enhancing the proposed MDRL algorithms by integrating complex decision-making, where agents can take actions ranging from mobility to deciding on the existence and reachability of the target, (4) developing a blockchain-based platform for Deep Reinforcement Learning as a Service (DRLaaS) allowing collaborative training and better accessibility to DRL solutions for target localization problems, and (5) ensuring the scalability of all the proposed solutions through the use of concepts such as Centralized-Learning and Distributed Execution (CLDE) MDRL methods coupled with Convolutional Neural Networks (CNNs) for optimized analysis of the agents' collected observations. Besides these contributions, we present several experimental studies and simulations that validate the proposed methods and compare them against existing state-of-the-art benchmarks in the literature
A Performance Enhanced 1-bit Bandpass Sturdy-MASH Delta-Sigma Modulator for Radio-over-Fiber Fronthaul Transmission Systems
Get PDFCurrently, in fronthaul transmission systems of radio access network (RAN), the digitized common public radio interface (CPRI) are being widely used. However, with the development of 5G technology, there are higher demands on signal bandwidth, transmission rate, energy consumption, and other aspects in the fronthaul system. Due to its limited spectrum utilization, high complexity, and large power consumption of the remote radio head (RRH), the digitized CPRI is difficult to satisfy the new requirements of 5G for fronthaul. As an efficient and concise modulation scheme, delta sigma modulator (DSM) can replace multi-bit analog to digital converters (ADCs) with the passive filters at the receiver side of the fronthaul, significantly reducing the complexity of RRHs while meeting the requirements for transmission rate and efficiency. Therefore, DSM has received widespread attention in recent years, and there have been significant developments in new technologies and various structures related to DSM.
In this thesis, an in-depth analysis and comparison of several different structures of 1-bit DSMs are conducted, and an enhanced 1-bit sturdy multi-stage noise-shaping (SMASH) structure for digital fronthaul systems is proposed. The proposed SMASH DSM is based on the traditional SMASH structure, with some structural changes and simplifications made to enhance the modulator's noise shaping capability. In the thesis, a 100-MHz bandwidth 64 quadrature amplitude modulation (64-QAM) orthogonal frequency division multiplexing (OFDM) signal is used as a digital baseband signal and then modulated onto RF carriers with a center frequency of 2.5 GHz. Finally, the bandpass 1-bit DSM modulation is performed. The process is simulated and experimentally verified.
The detailed comparison among the traditional single-stage delta-sigma modulator (SDSM), the traditional MASH and SMASH, and the proposed SMASH is presented in the fronthaul transmission system. The OFDM signal at radio frequency (RF) is quantized to two bits by SDSM/MASH/SMASH ADC, and this digitized signal is transmitted over 20-km single mode fiber (SMF) in a 2-level amplitude non-return-to-zero (NRZ) or 4-level pulse amplitude modulation (PAM4) intensity modulation direct detection (IM-DD) system.
Firstly, it is found that the proposed SMASH DSM has the widest input dynamic range (DR), which means it has better stability, followed by traditional SMASH and MASH, while SDSM performs the worst in terms of input DR.
Then, in the case of fiber transmission systems, the proposed SMASH has better noise suppression performance than the traditional MASH and SMASH schemes. In the directly received system, also known as the electric back-to-back (EBTB) system, the error vector magnitude (EVM) of the proposed SMASH, traditional MASH, and SMASH are -32.56 dB, -29.03 dB, and -29.31 dB, respectively. The proposed one has an around 3.2 dB EVM improvement. And in the IM-DD fiber transmission over 20-km SMF, the EVM are -24.71 dB, -23.93 dB, and -22.36 dB, respectively. The proposed scheme also has an around 2.4 dB EVM improvement compared to the traditional SMASH.
Finally, the comparison result of the two SMASH DSMs is verified in the experiment. In the case of EBTB and optical back-to-back (OBTB) systems, the proposed SMASH has an over 3 dB EVM improvement compared with the traditional SMASH. And in the case of the fiber link, the EVM for proposed SMASH is increased by 2.76 dB and 2.94 dB compared to the traditional one over the 8 km and 20 km fiber, respectively
Innovative Approaches for Real-Time Toxicity Detection in Social Media Using Deep Reinforcement Learning
Get PDFToxic comments on social media discourage user engagement and have serious consequences for mental health and social well-being. Such negativity heightens feelings of anxiety, depression, and social isolation among users, ultimately diminishing their experience on these platforms. For businesses, these toxic interactions are detrimental as they lead to reduced user engagement, subsequently affecting advertising revenue and market share.
Creating a safe and inclusive online environment is essential for business success and social responsibility. This requires real-time detection of toxic behavior through automated methods. However, many existing toxicity detectors focus mainly on accuracy, often neglecting important factors including throughput, computational costs, and the impact of false positives and negatives on user engagement. Additionally, these methods are evaluated in controlled experimental settings (offline tests), which do not reflect the complexities of large-scale social media environments. This limitation hinders their practical applicability in real-world scenarios.
This thesis addresses these limitations by introducing a Profit-driven Simulation (PDS) framework for evaluating the real-time performance of deep learning classifiers in complex social media settings. The PDS framework integrates performance, computational efficiency, and user engagement, revealing that optimal classifier selection depends on the toxicity level of the environment. High-throughput classifiers are most effective in low- and high-toxicity scenarios, while classifiers offering moderate accuracy and throughput excel in medium-toxicity contexts.
Additionally, the thesis tackles the challenge of imbalanced datasets by introducing a novel method for augmenting toxic text data. By applying Reinforcement Learning with Human Feedback (RLHF) and Proximal Policy Optimization (PPO), this method fine-tunes Large Language Models (LLMs) to generate diverse, semantically consistent toxic data. This approach enhances classifier robustness, particularly in detecting minority class instances.
The thesis also proposes a Proximal Policy Optimization-based Cascaded Inference System (PPO-CIS), which dynamically assigns classifiers based on performance and computational costs. This system improves efficiency by using high-throughput classifiers for initial filtering and more accurate classifiers for final decisions, reducing the workload on human moderators.
Extensive evaluations on datasets such as Kaggle-Jigsaw and ToxiGen demonstrate significant improvements in processing time, detection accuracy, and overall user satisfaction, contributing to the development of scalable, cost-effective toxicity detection systems for social media platforms
Parler français ou « bien » le parler? La langue et l’accent comme marqueurs de l’identité québécoise et les attitudes envers l’immigration
Get PDFQuelle place occupe la langue dans l’identité nationale québécoise? Depuis des siècles, la province francophone martèle l’importance de sa culture distincte au reste du Canada et la langue en est sa manifestation la plus saillante, voire la plus fondamentale. Pour cette raison, les politiques publiques qui visent à défendre la langue sont bien établies dans le paysage politique québécois. Ceci dit, on en sait très peu sur l’importance qu’accordent les membres du groupe majoritaire à cette caractéristique. Et force est d’admettre qu’on en sait encore moins au sujet de l’accent québécois, bien qu’il soit un marqueur d’altérité depuis que le Québec accueille un nombre grandissant d’immigrants à la suite de la Révolution tranquille. Ce mémoire démontre que l’accent québécois est une composante importante de la construction de l’identité nationale, mais que la langue française est bien plus importante dans cette construction. De plus, nous révélons que les ramifications de la langue et de l’accent sur les attitudes envers l’immigration sont différentes. Alors que celles de l’accent québécois sont exclusives, celles de la langue française sont parfois exclusives et d’autres fois inclusives, ce qui atteste de l’ambiguïté entourant le rôle de la langue dans la construction de l’identité nationale. La recherche repose sur un sondage réalisé au Québec en ligne en 2022 auprès de 2 401 personnes du groupe majoritaire (nés au Canada, non-membre d’une minorité visible ni autochtone, dont la langue maternelle est le français)