21793 research outputs found
Sort by
Exploring Songwriting as an Expression of Grief on Cape Breton Island: A Qualitative Interview Study
Get PDFLocated in Nova Scotia, Canada, Cape Breton Island has a rich and diverse musical history, which includes a well known tradition of using songwriting as a mourning practice. In the cultural mosaic of Canada, individuals seeking music therapy bring unique and diverse cultural backgrounds and experiences. Exploring these inherent cultural practices can offer models for integrating songwriting into the bereavement music therapy session while also informing music therapy training programs and future research. Grounded in a social constructivist epistemology (Hillier, 2016) and using qualitative semi-structured interviews, this study explored the use of songwriting as an expression of grief for three Cape Breton Island songwriters. Utilizing inductive thematic analysis (Braun & Clarke, 2006), three themes emerged from the data. The first theme, transformative grief journey, reflects how participants experienced personal growth while moving through a unique journey through their grief. Songwriting helped them process a range of emotional responses, provided a safe outlet for difficult feelings, and supported their ability to move forward. The second theme, cultivating connection, highlights how songwriting reduced feelings of isolation and helped participants feel connected to others through sharing their music. The third theme, intrinsic practice, describes how participants trusted the songwriting process and emphasized its cultural importance as an inherent and meaningful way to express grief
Artificial Intelligence for Spectrum-Aware Autonomous Wireless Networks
Get PDFSpectrum is one of the most vital public resources, carrying wireless communications for mobile phones, satellites, and emergency services. Traditionally, spectrum has been exclusively licensed, resulting in occasional underutilization. However, with the rapid proliferation of wireless devices, spectrum congestion has become inevitable, especially in unlicensed networks such as the Internet of Things (IoT), vehicular, and Unmanned Aerial Vehicles (UAVs), where devices rely on a limited number of public frequency bands that may not support large-scale communications. This contradiction between licensed spectrum underutilization and unlicensed spectrum congestion necessitates a rethinking of spectrum allocation and management strategies. In this thesis, we draw inspiration from Cognitive Radio (CR) technology, which equips radio devices with capabilities such as perception, reasoning, and judgment, and extend it to “intelligent radio” that integrates both cognition and learning capabilities. Our work advocates a shift from traditional model-driven approaches that rely on domain knowledge and strong assumptions to data-driven methods that learn directly from raw data and constant interactions with the environment. With the support of Artificial Intelligence (AI), we design intelligent spectrum borrowing and spectrum-sharing techniques that enable wireless devices to operate opportunistically on licensed bands. Specifically, this thesis explores how AI can endow wireless devices with context-awareness, self-optimization, and self-management capabilities for tasks such as dynamic spectrum access, power management, resource allocation, and ensuring security. Additionally, we develop solutions and frameworks for self-sustaining wireless devices that leverage Energy Harvesting (EH), bringing us closer to the realization of green networks. Our AI-driven algorithms are designed with computational efficiency in mind to minimize the burden on resource-constrained devices.
To drive context-aware intelligence in large-scale cooperative networks, we develop various unsupervised Machine Learning (ML) approaches for spectrum sensing. Unlike existing methods, the proposed frameworks operate without the need for labeled data, prior knowledge of the radio environment, or cooperation between licensed and unlicensed users. The approach ensures robust spectrum sensing while minimizing computational overhead for unlicensed users with limited capabilities. Moreover, we investigate how dimensionality reduction can improve computational efficiency and model generalizability. We expand the use of unsupervised learning to hybrid CR networks to allow devices to detect all licensed network states, opening up new opportunities for dynamic spectrum access.
To improve spectrum reasoning and analysis, we introduce some of the first fully unsupervised, data-efficient deep representation learning frameworks. These frameworks are designed to learn effective and disentangled representations of radio environment data. We demonstrate their effectiveness in significantly enhancing spectrum gap detection in small-scale cooperative networks. Additionally, we tackle key challenges of unsupervised learning, such as sensitivity to initialization and the need for predefined cluster counts. In large-scale networks, we propose a generative deep representation model that not only learns efficient representations but also captures the distribution of radio environment data, enabling the generation of new, unseen samples.
To facilitate edge intelligence and enhance the privacy of intelligent radios, we propose the first fully unsupervised deep Federated Learning (FL) framework for secure and distributed spectrum sensing in large-scale mobile networks. By leveraging user mobility across a large geographical area, the method enhances spatio-temporal diversity without requiring the transmission of private data to a central unit for processing. Instead, data is collected locally, and a shared model is collaboratively trained in a decentralized manner, significantly reducing communication overhead and safeguarding user privacy.
We tackle the growing challenge of spectrum scarcity in Cognitive IoT (CIoT) networks, where the demand for spectrum is increasing due to the expansion of connected devices. To address this, we develop intelligent and adaptive control algorithms for the joint management of network resources in spectrum-sharing environments. First, we formulate optimization problems under various constraints and model the decision-making process of a CIoT agent in the dynamic radio environment. We then propose two novel Deep Reinforcement Learning (DRL) algorithms that enable devices to autonomously learn operational strategies to optimize network resources and maximize long-term throughput without comprehensive prior knowledge. Additionally, we introduce innovative exploration strategies to enhance the CIoT agent's ability to identify optimal actions that maximize data rates. Considering the resource limitations of these networks, the algorithms are designed to be lightweight to reduce computational burdens on users. We also integrate EH techniques, such as Wireless Power Transfer (WPT) and Simultaneous Wireless Information and Power Transfer (SWIPT), to make these networks self-sustaining.
Finally, to develop dynamic strategies for navigating hostile spectrum-sharing environments impacted by jamming attacks, we propose an intelligent DRL approach that does not rely on frequency hopping. This algorithm is designed for rapid convergence, energy efficiency, and adaptability to adversarial conditions. We begin by formulating the optimization problem of power control under various constraints and modeling the decision-making process of the CIoT agent in such a hostile environment. Then, we introduce a novel interference-aware exploration strategy that enables the CIoT device to autonomously learn a transmission strategy, effectively mitigating jamming attacks and maximizing performance. Furthermore, we leverage WPT EH to allow the CIoT agent to convert jamming interference into a valuable resource for recharging.
In summary, the contributions of this thesis lay the foundation for a new generation of intelligent, autonomous wireless networks that are both spectrum-aware and agile, capable of optimizing resources and adapting to dynamic and complex environments
Development of Computation-Efficient Computer Vision Systems for High-Quality Brain Tumor Segmentation
Get PDFIn this thesis, two design methodologies are proposed, and also applied in the development of 2 computer-vision systems for computation-efficient and high-quality brain-tumor detection.
The first methodology aims at developing systems to detect, by conventional image processing procedures, 3D-object locations with a pixel-wise precision. The main operations of the detection are predicting gray-level distribution of the pixels in the object region and, based the prediction result, identifying/removing regions of non-interest. As 3D inputs can be sliced into axial, coronal or sagittal slice series, the prediction/identification/removal operations are performed step-by-step to the 3 series, respectively. Each removal increases the density of the object-information, facilitating the next prediction. To comprehend the pixel distributions with their locations, a 2D histogram presentation is proposed. In the design of the brain-tumor detection system, it is used to highlight the left-right asymmetry of a brain structure. Since the asymmetry is caused by tumors and non-pathological elements, an adaptive histogram modulation method is proposed to enhance the former by attenuating the latter. The prediction/identification/removal operations transform a 3D brain image into a tumoral minimum bounding box, which is then transformed into a tumor mask using simple morphological operations. The test results, on 1251 samples, have confirmed the high quality of the prediction of the tumor data distributions and the tumor detection.
The second methodology is proposed to design CNN (convolutional neural network) systems handling a complex task of brain-tumor segmentation, i.e., classifying the pixels of a brain image into 4 classes of intra-tumoral regions and the background. The methodology is to decompose this complex task into simple subtasks and each of them is performed by a simply-configurated and independently-trained CNN. By doing so, one can optimize the use of computing power and minimize the gradient conflict in training. The 4-class classification is decomposed into 3 binary classifications. Each of them is further decomposed into 2: first locating the object region and then identifying the pixels inside the region, performed by 2 independent-and-simple modules. The overall system, requiring only 0.75M trainable parameters, has been trained/tested with BraTS datasets, and its processing quality is among the best reported recently
Transaction Graph Analysis for Bitcoin Address Classification: Traditional Supervised Machine Learning and Deep Learning Methods
Get PDFIn this thesis, we consider the problem of Bitcoin address classification and clustering, common in the domains of law enforcement and regulatory compliance. We build a machine learning-based classification framework which is able to attribute a Bitcoin address to one of the predefined classes or to a specific company. We consider five distinct classes for coarse-grained classification: cryptocurrency exchanges, online marketplaces, mining pools, fundraising/charity platforms, and gambling; and 180 companies for fine-grained classification. Classes and the companies were selected so that they represent a broad spectrum of entities and activities within the Bitcoin ecosystem.
This thesis has three main contributions. First, due to the lack of publicly available datasets suitable for testing machine-learning classification algorithms, we create our own labeled dataset consisting of 3M Bitcoin addresses (from 2016-2022), with each Bitcoin address assigned a ready-to-use vector of carefully crafted features. Second, using this dataset, we conduct a comparative analysis of different machine-learning techniques and features for classification. Finally, we develop
two types of classifiers: based on the Boosted tree algorithm and the neural network-based classifier. Both are able to attribute a Bitcoin address to one of the predefined classes/companies. Our binary classification model achieves an F1 score of 76% using the Boosted tree algorithm, while our deep learning model achieves a 90% F1 score for multi-class classification with an accuracy of 92% and 28% higher than related work correspondingly. We achieve 67% accuracy for linking Bitcoin addresses to one of the 180 companies with our deep-learning model
Understanding the Lifecycle of Flaky Tests and Identifying Flaky Failures
Get PDFSoftware testing is a critical aspect of ensuring the quality of software. Ideally, tests should
produce consistent results when being executed repeatedly on the same version of the software.
However, certain tests may exhibit non-deterministic behavior, commonly known as ªflaky testsº.
These tests can provide ambiguous signals to developers and make the test results unreliable.
Despite being a recognized phenomenon for decades, academic attention towards test flakiness
has only recently increased. The current dissertation aims to contribute to the advancement of
research in two directions. First, we focus on predicting the lifetime of a flaky test, an issue that
has been left unaddressed in the flaky tests research area. Secondly, we question the efficiency of
previous studies in discerning flaky failures from legitimate failures, focusing on the Chromium
build result as our dataset.
In our investigation of the historical patterns of flaky tests in Chrome, we identified that 40% of
flaky tests remain unresolved, while 38% are typically addressed within the initial 15 days of introduction.
Subsequently, we developed a predictive model focused on identifying tests with quicker
resolutions. Our model demonstrated a precision of 73% and a Matthews Correlation Coefficient
(MCC) approaching 0.39 in forecasting the lifespan class of flaky tests.
Furthermore, we discovered that current vocabulary-based flaky test detection approaches misclassify
78% of legitimate failures as flaky failures when applied to the Chromium dataset. The
results also revealed that the source code of tests is not enough indicator for predicting flaky failures,
and other execution-related features must be contributed for better performance
Exploring Country-Level Factors Impacting Food Loss in the Global Food Supply Chain
Get PDFFood loss is a major problem throughout the world and there are many reasons why food loss occurs. This research investigates the impact of country level factors on food loss across the harvest, storage, processing, and transportation stages of the global food supply chain. A thorough literature review was carried out to examine various research papers on factors causing food loss and the most pertinent country level factors were chosen for the study. Food loss and country level factors data was obtained for several countries and spanning several years, and multiple regression analysis was carried out. The findings of this research and similar studies in future can have a significant impact on devising policies and fuelling investments to improve food supply chain infrastructure for reducing food loss across the world
Development of a Multi-Scale Model for Enhanced Performance by Using a Modified Soft-Margin SoftMax Loss and its Application for Facial Expression Recognition
Get PDFThis thesis introduces an improved Convolutional Neural Network (CNN) model that selectively excludes irrelevant elements from input vectors to extract distinct features. The model employs Convolution (Conv) to encode data information and Deconvolution (Deconv) layers to reconstruct the spatial dimensions of feature maps and utilizing shortcut connections to exploit applicable sparsity and capture comprehensive and detailed information. This research addresses the limitations of the traditional SoftMax Loss function, which tends to overfit by wrongly classifying due to shortage of discriminability through integrating a regularization technique to the conventional Cross-Entropy loss and introducing an adaptive-margin to the standard SoftMax function. The adjusted SoftMax Loss is designed to enhance the separation of different embedding vectors and tighten clusters of similar ones. This adjustment results in boosting both the diversity between different classes and the similarity within the same class. These modifications aim to elevate the proposed model’s accuracy, and evaluations are benchmarked against existing methods using well-known datasets such as CIFAR10 , MNIST, and SVHN. Lastly, the application of the model is fine-tunned in the domain of facial expression recognition (FER). This work asserts the model's advanced capabilities over state-of-the-art methods in FER and demonstrated its effectiveness through superior accuracy on three established datasets: FER-2013, RAF-DB, and CK+
Special Rational Solutions of the Fifth Painlevé Equation and their Asymptotic Behavior.
Get PDFIn this thesis, our first goal is to formulate a generating function and compute its moments alongside the corresponding Hankel determinant. When the latter is nonzero, we will prove that for Painlevé 5, we can construct a Lax pair whose solution is a combination of the solution of the Riemann Hilbert Problem (RHP) and the generating function. An ingredient of that solution, called the Hamiltonian will be used to construct the Tau function which solves the ODE Painlevé V. As such, it will be easy to show that when the Hankel determinant vanishes, the RHP is not solvable, and its zeroes correspond to the poles of the rational solution of the ODE Painlevé V i.e. the Tau function.
On the other hand, an asymptotic analysis will be conducted to prove that the domain of the poles of the rational solution of the ODE Painlevé V (its domain of non analyticity) defines a well shaped region with boundaries on the complex plane as the size of the square Hankel matrix goes to infinity
Machine Learning for Fault Prediction in Clouds
Get PDFThe vast adoption of cloud computing has increased the size and complexity of data centers, increasing possibility of faults. Fault can negatively impact the performance, availability, and reliability of cloud services, leading to significant maintenance cost and revenue loss for cloud service providers. Therefore, fault prediction in clouds is a critical task. Machine Learning (ML) is increasingly used for this purpose due to their pattern recognition capabilities. While predicting faults in clouds using ML enables a proactive approach to prevent faults, building accurate prediction models that can maintain their performance in dynamic clouds is challenging. One problem is concept drift, where changes in data distribution can degrade model performance. Similarly, feature drift, which is changes in feature relevancy, can also degrade the model performance. Additionally, models accuracy is influenced by data-related parameters, necessitating selection of these parameters to achieve a high model performance. Existing ML-based fault prediction solutions do not focus on adaptability to dynamic conditions like concept or feature drift. Additionally, selecting data-related parameters to balance model performance and resource consumption is not addressed in current literature.
This thesis mainly focuses on addressing the challenges of employing ML models for predicting faults and predicting application performance degradation caused by faults in cloud environments. We first propose a concept drift adaptation algorithm for fault prediction in clouds using Reinforcement Learning (RL). This algorithm considers the cloud operator's requirements, and uses RL to select the most appropriate drift adaptation method as well as data size for adaptation that fulfills the requirements. Second, we propose a feature drift adaptation solution for adapting the model to feature drifts while predicting application performance degradation in clouds. This solution consists of a feature drift detector that monitors the performance of the prediction model as well as the feature importance, and a feature drift adaptor that measures the drift severity to adapt the prediction model. Finally, we propose a multi-objective optimization algorithm to select the training data size, data sampling interval, input window, and prediction horizon for training an ML model that predicts application performance degradation in clouds
Shopping Centre Operations Amid the Digital Divide: The Role of Anchor Stores
Get PDFThis study investigates how anchor store clusters within shopping centres interact with internet inequality and socio-economic inequality to influence foot traffic. Foot traffic data from over 400 shopping centres owned by Regency Centres spanning 2017 to 2019 are analysed. We use Principal Component Analysis (PCA) to explore the configuration of anchor store clusters. Key findings include: 1) Higher internet inequality correlates with increased foot traffic, suggesting a preference for in-store shopping, while overall socio-economic disadvantages have minimal impact on overall foot traffic. 2) Specific anchor store clusters, such as diverse non-essential, diverse essential, budget-mix, grocery, and home clusters, significantly enhance foot traffic. In contrast, clusters focusing on health and personal care weaken the effect due to their narrow-range nature. 3) Internet inequality and anchor store types create complex behaviour patterns, with increased reliance on diverse hardline, essential goods, budget-mix, and grocery and home improvement clusters in areas with limited internet access. 4) Some anchor store clusters are significantly beneficial for socio-economically disadvantaged populations facing internet access challenges, emphasizing the value of accessing a broad range of goods in one location, particularly in diverse non-essential and essential general merchandise stores, along with budget-mix and healthcare-oriented stores. These findings suggest that strategic anchor store placement as a means to mitigate the impacts of digital disparities and enhance shopping centre performance, underscoring the importance of inclusive retail environments.
Key Words: Shopping centres, anchor stores, foot traffic, digital divide, internet inequality, socio-economic disparitie