21793 research outputs found
Sort by
Adoption of Deep Learning Models and its Applications in Dementia Research
Get PDFArtificial Intelligence (AI) is at the forefront of the Fourth Industrial Revolution, fundamentally transforming industries and societies through unprecedented automation and data-driven applications. The Fourth Industrial Revolution is characterized by a fusion of software and hardware improvements, creating a seamless blend of the physical, digital, and biological spheres. These improvements make it possible for AI to leverage and process vast amounts of information to generate actionable insights, and perform complex tasks more quickly and more accurately than humans, leading to more informed decisions and efficient processes.
Despite its success and promising results in other domains, the adoption and integration of AI innovations in healthcare has been complex and slow. Few AI innovations have met with success and have been incorporated into daily practice. This thesis addresses technological, legal and ethical issues that must be mitigated before AI-based systems can be fully adopted and trusted into clinical trials and workflows. We identify an opportunity to further the state-of-the-art of AI solutions and their adoption in healthcare through privacy-preserving aggregation algorithms and human-centered evaluations of transparency in clinical decision support systems.
In particular, this dissertation explores advanced methodologies in Federated Learning (FL) for improving collaborative learning, data privacy, and decision-making across various domains. We improve the core FL aggregation algorithm for better handling the learning of distributed heterogeneous data sources, with a method named Precision-weighted Federated Learning. We perform extensive evaluations with benchmark datasets on resource-constrained environments to measure its limits and perform additional tests on clinical data to enhance the quality of clinical assessment analysis, validating its utility.
Our research also aims to understand how to visualize AI model outputs to enhance transparency in clinical decision support systems. We conduct extensive evaluations to assess the impact of visualizing AI uncertainty and personal traits on decision-making, supporting the design of AI outputs that are interpretable by clinicians. Initially, we explore the effects in low-risk gaming scenarios, followed by an examination of AI uncertainty representation in high-stakes clinical decision-making, particularly in Alzheimer’s disease prognosis.
In summary, this dissertation presents significant advancements in FL and clinical decision support systems. We address some of the current limitations and challenges of adopting AI systems, and demonstrate improvements in collaborative learning, data privacy, and human-AI decision-making. These findings offer valuable insights for designing robust, efficient, and trustworthy AI and FL systems. We believe that user-centered design practices will eventually play a more prominent role in the development of AI tools and technologies, becoming the driving force behind moving innovations from the laboratory to the clinic
Rewilding second language learning: Non-formal learning through songs
Get PDFThe goal of this manuscript-based dissertation is to explore French second language development through songs in the non-formal learning space to understand learners’ progression towards successful integration into informal environments. To explore the self-regulatory process required to increase autonomy and the degree to which music can aid learners in this pursuit, I examine new technologies and learners’ behaviours towards the use of music in second language contexts. This is done by following Cardoso’s (2022) chronological framework for examining new technologies in second language contexts because when learners engage with L2 music, they do so through technology. After providing an introduction and background on autonomy and the non-formal learning space (Chapter 1), the framework is explored through three manuscripts to investigate different aspects of language learning through songs and the use of the Bande à Part app (which was used as a tool to understand this approach).
The first manuscript (Manuscript A) investigates theoretical affordances provided by music as a form of L2 phonological exposure (e.g., acoustic comparisons between singing vs speaking) and form of content (e.g., type of vocabulary, rhyme schemes, and repetition). This manuscript highlights affordances provided by overlapping cognitive processes shared between language and music processing. For example, music encourages increased focus on form, which could help with pronunciation and exposure to drawn out vowels, leading to better formation of vowel categories. In terms of content, music is repetitive and motivating—two factors that are difficult to incorporate into the classroom. Empirical work on the impact of music on language development is interpreted with this in mind, supporting the theoretical affordances that were highlighted. Gaps are highlighted for future studies with the subsequent two chapters shedding light on some of these gaps.
The second manuscript (Manuscript B) assesses learners’ perceptions of suitability and acceptability of learning French with the Bande à Part music app (development and release documented in Sundberg & Cardoso, 2019). Results demonstrate that learners believe the app helps improve their French and that it is relatively intuitive to use, although more instructions and more songs should be included. These perceptions are gathered because they are predictors of whether or not learners will use a tool, critical to self-directed contexts. The feedback gathered is also used to make iterative improvements to the app. The manuscript discusses principles and takeaways for future material developers who are interested in student-centred tools.
The third manuscript (Manuscript C) investigates learners’ awareness of language, the language learning process, and ability to self-regulate their own pronunciation development. This is accomplished through think-aloud exercises where participants listen to and then comment on recordings of themselves imitating sung and spoken audio recordings. This research helps to illuminate where scaffolding is needed and the degree to which learners can learn on their own.
The main findings emphasize the value of self-directed pedagogical tools, particularly songs, in enhancing L2 development. Educators and developers can use these insights to create more effective materials for language learners and support beyond the classroom learning endeavours
The effects of diet and probiotic administration on lipid dysregulation and gut microbiota dysbiosis during atherosclerosis development
No full textAtherosclerosis is an inflammatory disease caused by lipids, cells, and debris accumulating in arterial walls. Despite their popularity for effective weight loss, low-carbohydrate/high-protein diets (LCHP) are more atherogenic than Western diets and are associated with adverse long-term cardiovascular outcomes. However, the mechanisms of this atherogenicity are not currently known. Therefore, the objectives of this work were to determine if circulating lipid profiles, inflammatory lipid mediators, and/or gut dysbiosis contributed to the higher atherogenicity of the LCHP diet in ApoE KO mice during the early development of plaques. Secondly, it was hypothesized that supplementation of two probiotics: Lactobacillus helveticus Lafti 10® and Bifidobacterium bifidum Rosell 71® could reduce diet-induced atherogenicity during the early stages of the disease due to their antilipemic, anti-inflammatory and antioxidant properties.
In order to accurately observe the changes in lipid and inflammatory pathways over a 6-week longitudinal study in mice, an optimization of analytical methods was necessary to ensure appropriate lipid stability during analytical handling and compatibility with low blood volumes (30-50 μL). First, the impacts of short exposure to room temperatures and the effect of the addition of a lipase inhibitor, Orlistat, during lipidomics sample preparation were investigated using liquid chromatography-mass spectrometry analysis. Human plasma samples were prepared using isopropanol protein precipitation at room temperature, 4°C, and -80°C with, or without, Orlistat. Subtle changes in the lipidome were found even for short (<2 hr) exposure to room temperature, so maintaining a strictly controlled temperature of 4°C throughout the sample preparation process is recommended to reduce residual enzymatic activity. Room temperature exposure during sample preparation impacted the levels of 19% of the lipids but the addition of Orlistat reduced this impact to less than 2.4% of measured lipids and improved method precision across all temperatures. To enable oxylipin determination from 15 μL of plasma, solid-phase extraction – liquid chromatography – high-resolution mass spectrometry method was further optimized. This optimization maintained good coverage, acceptable recovery (71 to 100%) and excellent precision (RSD < 11%), thus making the method suitable for longitudinal monitoring of oxylipin status in mice using small blood volumes collected by tail bleeding.
Next, the effects of diet and probiotic supplementation on lipid and gut dysbiosis were investigated. ApoE knockout mice were fed LCHP, W, or standard chow diet for 6 weeks, with or without anti-inflammatory probiotic supplementation (Lactobacillus helveticus Lafti 10® and Bifidobacterium bifidum Rosell 71®) at low- or high-dose. Targeted and untargeted analyses using liquid chromatography – high-resolution mass spectrometry was used to study dyslipidemia and inflammatory pathways, whereas gut microbiome analysis was performed using 16S rRNA sequencing. When comparing the effects of the diet, the time course study showed that the microbiota in the fecal pellets of the mice given W and LCHP diets were globally similar. However, gut dysbiosis occurred faster for the mice fed the LCHP diet compared to the W diet. In addition, mice fed the LCHP diet had a higher abundance of deleterious families such as Sutterellaceae Parasutterella and beneficial families such as Muribaculaceae Muribaculum, Bacteroidaceae Bacteroides or Erysipelotrichaceae Dubosiella than the W diet. Furthermore, the relative abundance of beneficial species such as Bifidobacteriaceae and Eurobacterium Coprostanoligenes and deleterious families such as Erysipelotrichaceae Turicibacter, Bifidobacterium, Streptococcaceae Streptococcus, Lachnospiraceae, and Oscillospiraceae were decreased compared to the W diet groups. Plasma levels of trimethylamine N-oxide (TMAO), a pro-atherogenic gut-derived metabolite, were unchanged between the LCHP and W groups, indicating that the increased atherogenicity of the LCHP diet is not promoted by this pathway. In terms of lipid dysregulation associated with early stages of atherosclerosis, LCHP and W diets showed similar oxylipin profiles but distinct from the C diet. When comparing the LCHP and W diets, only eicosapentaenoic acid (EPA), which prevents atherosclerotic plaque formation and stabilizes it, was lower in the plasma of the mice fed the LCHP diet compared to the W diet. Circulating oxylipin plasma profiles from mice fed LCHP and W diets showed a similar increase of 19,20-dihydroxy-4Z,7Z,10Z,13Z,16Z-docosapentaenoic acid (19,20-DiHDPA) and a decrease of arachidonic acid (AA) with time. In addition, 8S,15S-dihydroxy-5Z,9E,11Z,13E-eicosatetraenoic acid (8,15-DiHETE) and 15-hydroxy-5Z,8Z,11Z,13E,17Z-eicosapentaenoic acid (15-HEPE) increased with the LCHP diet over time. In contrast, only 5,6-dihydroxy-8Z,11Z,14Z,17Z-eicosatetraenoic acid (5,6-DiHETE) increased with time with the W diet. Overall, the inflammatory pathways were similar in both LCHP and W diets and are unlikely the cause of the higher atherogenicity of the LCHP diet. In addition, the 9-HODE/13-HODE ratio increased with time for the LCHP and W diets indicating a pro-inflammatory profile for both diets compared to the C diet. Plasma levels of sphingolipids, lysophospholipids, phospholipids, alkyl-phospholipids, cholesterol, and cholesteryl esters levels were lower in the LCHP diet compared to the W diet but higher when compared to the C diet. In contrast, glycerolipids were higher in the LCHP diet compared to the W diet.
The time-course study revealed that probiotics had minimal effect on the oxylipin profiles and the 9-HODE/13-HODE ratio was not affected by the probiotic supplementation demonstrating limited anti-inflammatory properties. However, the administered probiotics showed antilipemic properties by reducing highly unsaturated lipids from phosphatidylethanolamine and triglyceride classes. In addition, alkyl phospholipids were elevated with the probiotic administration in combination with the LCHP diet. Furthermore, probiotic supplementation also positively influenced the gut microbiota composition by increasing beneficial families such as Tannerellaceae and Clostridia Vadin BB60 genera, and decreasing harmful bacteria including Streptococcaceae, Erysipelotrichaceae, Lachnospiraceae, and Atopobiaceae. Overall, the effects of probiotic supplementation showed a strong dependence on the diet type indicating strong diet-microbiota interaction. In conclusion, this thesis establishes for the first time systematic changes in circulating lipid levels and gut dysbiosis that are contributing to the atherogenicity of the LCHP diet and examines how probiotic supplementation impacts these adverse changes. These findings are critically important to ensure highly popular weight loss interventions such as LCHP diets do not cause unintended adverse cardiovascular consequences
Microstructural, Mechanical and Tribological Investigation of Copper-Based Coatings for Extreme Environments.
Get PDFThermal spray coating methods such as Atmospheric Plasma Spray (APS), High-Velocity Oxygen Fuel (HVOF), and newly developed High-Velocity Air Fuel (HVAF) offer one of the best technique to effectively protect new parts from high temperature, wear, corrosion, and residual stresses as well as produce hard and dense coatings which in turn helps in the improvement of the lifespan of the material. These techniques provide coatings that enable the enhancement and prolongment of component life and the reduction of component cost due to the improvement in the functionality of a low-cost material. There is also an opportunity to revamp worn parts to their original dimensions, without the need to replace the entire component. However, traditional thermal spray techniques tend to produce coatings with high oxide content and porosity resulting in undesirable coating properties. On the other hand, Cold Spray (CS), a more recent coating technique, can produce highly dense coatings with strong bonding through the plastic deformation of powder feedstock, thus avoiding most of the aforementioned issues.
The main purpose of this research is to study the tribological properties of copper-based coatings deposited by APS, HVAF and Cold spray (i.e. from high temperature to low temperature deposition methods) to determine their suitability for extreme environments. This thesis is comprised of two research studies, The first study emphasizes on development of APS-CuNi coatings on different substrate materials to investigate the influence of the substrate materials as well as the tribological performance of the coatings at various temperatures (room and elevated temperature). The second study focuses on the evaluation of copper coatings by low temperature deposition systems (HVAF and CS) to understand their microstructural, mechanical and tribological behavior (room and elevated temperature).
The microstructural evaluation of coatings produced by both studies was analysed using the scanning electron microscope (SEM). Also, the Vickers microhardness tester was used to measure the coatings hardness in both studies. Furthermore, the tribological performance of coatings developed in both studies was conducted with a reciprocating tribometer using a ball-on-flat configuration and wear profiles were measured using confocal laser microscopy. Ex-situ characterization of the worn coatings was performed using a scanning electron microscope (SEM), energy-dispersive X-ray spectroscopy (EDS), and Raman spectroscopy. In the first study, APS coatings showed lower wear rates at room temperature in comparison to high temperature sliding. With regards to the second study, cold sprayed copper coatings showed a better wear resistance compared to the HVAF copper coatings
Public Security Institutional Voices and Discursive Frames on Police Killings in Brazil
Get PDFOver the years, the military police have been a significant perpetrator of violence in Brazil, targeting mainly Afro-Brazilians in favelas and poor suburbs. This thesis investigates the case of the military police of Rio de Janeiro to understand how, over time, public security institutions have framed discourses about the use of lethal force in Black and low-income territories. It applies the framework of discursive institutionalism, the methodology of discourse analysis and a comparative approach to analyze emblematic episodes of lethal police violence that occurred across four political eras. The results suggest that over time discourses have operated more to sustain than to change the way public security institutions deal with police killings. This was expressed by: a) arguing that the contexts of intervention compelled the police to use lethal force; b) asserting that the institution was not directly responsible for the killings; c) explicitly defending the public security model in force or the police approach in the episodes analyzed; d) claiming that there was a political interest behind the repercussions of these episodes. When public security institutions, in addition to perpetuating hyper-violent approaches, fail to offer the public a more critical discursive framing of this problem, they further contribute to the trivialization of Afro-Brazilian deaths in low-income territories. This work demonstrates pathways for understanding how policy institutions create argumentative and ideological mechanisms to justify policy choices. It highlights the significance of discursive institutionalism in comprehending the role of discourse not only in processes of policy change but also in situations of policy maintenance
Exploring How AI Disclosure in Blog Posts Affects the Perceptions of Brand Warmth and Competence
Get PDFMarketers increasingly employ generative AI technologies in content creation. However, whether and how disclosing AI authorship may affect consumers’ brand perceptions remains underexplored. This thesis investigates the impact of disclosing whether marketing content was AI-generated or human-written, specifically on consumer perceptions of the brand’s warmth and competence. In general, I hypothesize that brand perceptions will be more negative when content (e.g., a blog post) is disclosed to be authored by AI compared to by a human. However, I theorize the effect may depend on whether the content is informational or narrative in nature, such that the negative impact of AI disclosure is expected to be greater for narrative content than informational content. These hypotheses were examined across two experiments. The results of Study 1 revealed that human-written content was perceived as warmer and more competent, enhancing brand credibility, brand attitude, and purchase intentions. Content type did not moderate these effects. Study 2 replicated the effect of disclosure on warmth, and additionally found that content type moderated this effect (although the moderation did not operate as expected). No significant effects on competence emerged in this study. Theoretical and managerial contributions are discussed, especially regarding the strategic use of human authorship in content marketing
Integrated Optimal Design and Operation of Compressed Air Energy Storage for Decentralized Applications
Get PDFThis thesis aims to investigate the integration of compressed air energy storage (CAES) technology into decentralized energy systems, addressing associated technological and integration challenges within the dynamic energy system environment. A multi-layer simulation-optimization framework is developed to comprehensively evaluate the feasibility of integrating decentralized CAES into local hybrid energy systems (HES) through optimal sizing and operation. In the first layer, an improved energy management operation strategy (I-EMOS) is designed to enhance the integration of adiabatic-CAES (A-CAES) systems into decentralized applications. In doing so, the interaction and limitations of A-CAES subsystems, including power conversion units, air storage tank, and thermal energy storage, are considered to evaluate the long-term performance and dynamic behavior of A-CAES systems, especially when connected to intermittent renewable energy sources and end-user load demand. Subsequently, the second layer develops a holistic sizing-planning framework, including a generic A-CAES model and various alternative power dispatch strategies (PDS), based on the application potentials of A-CAES. This module aims to enhance A-CAES contribution while minimizing the levelized cost of energy and achieving the optimal configuration for the corresponding applications. Eventually, the final layer focuses on improving the resilience of the energy system, incorporating A-CAES technology, within scenarios involving limited energy sources and hybrid energy storage solutions. Therefore, an operational unit-commitment optimization model is developed, considering the A-CAES system's response and charging-discharging transition times. This model is integrated into the sizing-planning module to co-optimize the economic performance and system resilience through two-stage optimization, involving long-term planning and short-term scheduling. The methodology is applied to Concordia University buildings in Montreal, Canada. Validation against data from an existing A-CAES pilot plant shows a 42.5% improvement using I-EMOS compared to traditional EMOS.
Optimal configurations under various PDSs demonstrate energy cost savings between 0.021 per kWh, with significant improvements in electrical load management (52%) and carbon emission reduction (65%) for the system in which A-CAES is planned for both solar energy integration and seasonal load shifting. Furthermore, under the worst-case scenario (zero selling back), the HES achieves a PV self-consumption rate of around 92% and a payback time of 15.5 years. In scenarios of limited grid dependency, a substantial annual resiliency improvement of approximately 41.1% is achieved by integrating the energy storage system. Additionally, despite the superior cost performance of the PV/A-CAES system, the PV-based HES featuring hybrid A-CAES, and battery storage achieves a 47.3% electrical load management ratio and a 96% self-consumption rate, improving by about 6% over HES with only A-CAES system. Furthermore, findings indicate that under optimal operational conditions, even with the highest PV power availability during grid interruptions, the HES could meet 94% of load demand using individual A-CAES, increasing to 100% by integrating fast-response batteries. In conclusion, the proposed framework offers a reliable approach for integrating and customizing decentralized A-CAES systems, considering specific service requirements and constraints. It identifies critical times of loss of power probability, enhances understanding of local energy system design, and facilitates better integration with renewable energy sources and storage systems. The findings provide valuable insights for decision-makers, helping select suitable systems and scenarios based on key performance indicators. The framework also is adaptable to various scale scenarios, accommodating both local and regional generation considerations
Visual-Infrared Aerial Image Based Wildfire Intelligent Perception
Get PDFThis work focuses on the increasingly serious and urgent environmental problem of
wildfire, studying and testing the possible schemes, strategies and solutions in real
application of autonomous wildfire perception and fighting.
To efficiently tackle the wildfire fighting challenges of early detection and fast response
with unmanned aerial vehicles (UAVs), several intelligent computer vision algorithms are
studied, updated, and fine-tuned in this work. These algorithms are designed to work in
conjunction with UAV motion planning and path planning algorithms to detect (early)
wildfire spots based on aerial images, estimate the distance between wildfire spots and
UAV(s), geographically locate these wildfire spots and efficiently approach these spots for
firefighting.
The main contribution of this work is the design of an intelligent wildfire perception
system which utilizes both visible and infrared (VI) aerial image information, integrates
deep-learning (DL) filters for oriented features from accelerated segment test (FAST) and
rotated binary robust independent elementary features (ORB features) to geo-positioning
these wildfire spots. This work proposes a novel concept of combining DL models and ORB
based simultaneous localization and mapping (SLAM) technologies as UAV applications in
vision-based wildfire detection, estimation, and geo-positioning and management/fighting.
There are three main functional aspects of the visual-infrared image based intelligent
wildfire perception system in this work.
The first main aspect is wildfire detection, which includes wildfire image classification,
wildfire semantic segmentation, and wildfire spot(s) detection (object detection). For
wildfire image classification, an optimized ResNet-based neural network model is utilized to
achieve higher classification accuracy. For wildfire semantic segmentation, this work focuses
on the U-shaped deep network models (UNets) and proposes the application of original
UNet, an attention gate-enhanced UNet, and a SqueezeNet lightweight attention gate UNet
for early wildfire smoke and flame segmentation. For online wildfire object detection, the
model of you only look once version 5 (YOLOv5) and updated model of you only look
once version 8 (YOLOv8) are utilized to obtain accurate bounding boxes of wildfire spots,
YOLOv8 model can avoid pre-anchors and straightforwardly detect and track the center of
wildfire spots.
The second main aspect is the work of achieving UAV-wildfire distance estimation and
wildfire geo-positioning through monocular ORB-SLAM technology (SLAM2 and SLAM3).
This aspect has two main designs: The first one designs an attention gate UNet to filter ORB
feature points for wildfire distance estimation, achieves more robust results and detailed segmentation at the edges of wildfire smoke and flame spots. This design can be deployed
on ground workstations for detailed missions. The other one designs YOLOv8 filtering
ORB-SLAM3 features for online wildfire distance estimation and geo-positioning, which
can be deployed on the onboard computers for real-time wildfire spot recognition and geopositioning.
This lightweight and fast-responding application can combine with UAV path
and motion planning for online firefighting.
The third aspect comprises several smaller functions to achieve the wildfire perception
system integration. Most of these functions use the infrared information, because the
energy radiation information could support the deep learning wildfire detection to have
more confident and robust detection results. A geometry-based visible and infrared image
alignment and registration scheme is designed in this work. The image registration work
is the basis of the visible and infrared image fusion. Infrared images can also be used to
estimate wildfire spot temperature to guide the safe flight of UAVs. After that, there is a
design of online water retardant release mechanism, and it is briefly discribed
Real-Time Neural Cloth Deformation using a Compact Latent Space and a Latent Vector Predictor
Get PDFWe propose a method for real-time cloth deformation using neural networks, especially for draping a garment on a human body. The computational overhead of most of the existing learning methods for cloth deformation often limits their use in interactive applications. Employing a two-stage training process, our method predicts garment deformations in real-time.
In the first stage, a graph neural network extracts cloth vertex features which are compressed into a latent vector with a mesh convolution network. We then decode the latent vector to blend shape weights, which are fed to a trainable blend shape module. In the second stage, we freeze the latent extraction and train a latent predictor network. The predictor uses a subset of the inputs from the first stage, ensuring that inputs are restricted to those which are readily available in a typical game engine. Then, during inference, the latent predictor predicts the compacted latent which is processed by the decoder and blend shape networks from the first stage.
Our experiments demonstrate that our method effectively balances computational efficiency and realistic cloth deformation, making it suitable for real-time use in applications such as games
Coin Detection and Classification using a Few-Shot Learning method based on Siamese Network
Get PDFCoins are used in our daily lives for a long time with less depreciation than paper currency. Detecting counterfeit coins visually is a challenging way with lots of errors. This thesis investigates advanced machine-learning techniques to differentiate between counterfeit and genuine coins with a small dataset. It focuses on the implementation of few-shot learning. This study is applied to two different types of datasets. The first dataset contains the images converted to grayscale, and the second dataset contains the four slopes images. As the detection of counterfeit coins is challenging due to their high similarity with genuine coins, more features are required before pre-training the neural network.
For this study, 2,474 labeled images from the CENPARMI dataset belonging to 22 different classes were used. To enable experimentation, the dataset was split into two parts: a Main Dataset (Dm) and a Target Dataset (Dt). We used a pre-trained model, which learns from the Dm and adapted it to Dt. The Inception V3 network was fine-tuned in the main dataset to learn general coin characteristics. This knowledge was transferred to the target dataset to learn new coin types from a few images. FSL using Siamese networks and contrastive loss was used. The algorithm performance was evaluated using the total accuracy with different epochs and different batch sizes to earn the optimum of them, and also the precision and recall and F.score per class.
It is shown that the accuracy of our method in epoch 20 is optimal. At this point, the model achieves a high level of accuracy (92.13% for grayscale images and 94.73% for SMMIG images). the model trained with a batch size of 32 achieves the highest accuracy of 92.13% for the grayscale dataset and 94.73% for the SMMIG dataset, indicating that moderate batch sizes contribute to optimal performance