21793 research outputs found
Sort by
Performance Evaluation of the Object Detection Algorithms on Embedded Devices
Get PDFEdge computing has seen a dramatic rise in demand, driven by the necessity for real-time, low-latency applications across various domains from autonomous vehicles to surveillance systems. Among these, real-time object detection stands as a crucial technology. However, the inherent constraints of edge devices, including limited computational power, present significant challenges.
This thesis provides a comprehensive evaluation of several Convolutional Neural Networks based object detection models when deployed on resource-constrained edge devices, specifically Raspberry Pi and Google’s Coral TPU. The models examined include EfficientDet, YOLO, and variants of the MobileNet family combined
with SSD for object detection tasks.
We developed a novel benchmarking framework that allowed the evaluation of these models under different configurations, enabling an accurate assessment of their performance characteristics. The benchmarking framework and the metrics used for evaluation can provide a foundation for future work, focusing on the design and deployment of efficient real-time object detection models on edge devices.
The performance of these models was scrutinized based on an exhaustive set of metrics including processing speed (frames per second), model accuracy (F1 score), energy consumption, CPU utilization, memory footprint, and device temperature.
A novel benchmarking framework was developed to evaluate these models under diverse configurations, providing a precise assessment of their respective performance characteristics. This benchmarking framework, along with the evaluation metrics, sets the foundation for future research concentrating on the design and deployment of efficient real-time object detection models on edge devices. The findings of this study underscore the fact that no single model is a universal solution for all edge applications; instead, the choice of model is heavily dependent on the specific requirements and constraints of the given application. By offering a detailed overview of the performance traits of each model, we aim to guide practitioners in making informed decisions when deploying object detection models in edge computing environments. This work sets the stage for future exploration in the development of more efficient and effective models for real-time object detection on edge device
Refashioning the industry: Exploring the transition of women’s apparel brands to the circular economy in Montreal
Get PDFOver the last few decades, the fashion industry has evolved from providing well-conceived haute couture collections to the production of trendy, mass-produced, low-priced apparel, with reduced lead times. This consumption of ‘throwaway fashion’, at an alarming rate, has yielded extensive environmental and social detriments, thereby warranting industry reform and accountability. Driven by the tenets of sustainable development, brands within the fashion industry have since advanced to incorporate values that balance profits, people, and planet. As the circular economy gains traction, globally, to tackle sustainability challenges across industries, this research conducts a case study to explore the implementation of circularity strategies within the sustainable fashion industry in Montreal. Guided by Québec Circulaire’s Circular Economy Framework, practices employed by 90 sustainable apparel brands in Montreal, are identified and analysed to ascertain alignment with different circularity orientations and strategies. We found that a large number of sustainable brands in Montreal adopt practices that orient towards rethinking the extraction and use of virgin resources rather than optimising the use of extracted resources. They do so by employing a mix of strategies including eco-design, responsible procurement and consumption, and operations improvement. Further in-depth interviews conducted with 12 brands also revealed contextual factors that influence the transition of the industry to a CE model, including local culture and community, industrial and supply chain network, access to raw materials and skilled workforce, support from the government, and the cost of living in Montreal
Optimizing Reinforcement Learning: Fog and Edge Resource Management Through Bootstrapping and Reward Shaping
Get PDFThe rapid and extensive use of technology is unprecedented. From small devices like sensors and mobile phones to large systems like servers and data centers, a wide range of computing setups exists to meet human needs. However, the increased demand has raised concerns about whether these setups can handle the load. Fog and edge computing are concepts that bring servers closer to users to improve response time and service quality. But the availability of these fog devices is limited, highlighting the need for systems to manage computing resources. These systems' main task is to efficiently distribute services across available resources and adapt to changing needs. Existing resource management solutions still possess challenges and limitations with regards to the quality of decisions due to their increasing complexity.
In this thesis, our main motivation is leveraging AI in addressing resource management, which stems from its ability to intelligently handle complex and dynamic scenarios, such as optimizing service placement, predicting demands, and adapting to changing environments. AI's capacity to learn from data and make informed decisions offers a promising approach to efficiently manage computing resources in a rapidly evolving technological landscape. This research is motivated by four main goals: (1) creating a strong computing architecture that can meet diverse user needs across various applications managed by a resource management system; (2) using AI to develop resource management solutions that handle decisions like placement and scaling, as well as predict user demands and resource availability; (3) ensuring the AI solution is reliable despite potential errors by improving its performance or having a backup plan; (4) making the AI solution adaptable to sudden environmental changes to keep decisions effective.
The thesis aims to address these gaps by: (1) designing an effective networking and computing architecture in the context of on-demand fog and edge formation, while supporting an Intelligent Computing Resource Management solution (ICRM) for multi-types of applications through offline learning and bootstrapping; (2) using DRL to build the ICRM, driven by a Markov Decision Process (MDP) environment design that produces actions related to host selection and service placement while accounting for the change in user demands; (3) enhancing the proposed MDP by adding the support for predicting the change in both user demands and available computing resources, where the agent becomes capable for issues horizontal and vertical resource scaling decisions in multi-applications setting; (4) introducing the first solution to speed the learning speed of DRL agent by devising a Graph Convolutional Network solution as a potential-based reward shaping solution; (5) developing another reward shaping solution based on Convolutional Neural Network (CNN) carefully designed and inspired by the value iteration network (VIN), to speed learning. Besides these contributions, we present a set of experimental studies and simulations using real-world test cases for each of the contributions compared to state-of-the-art solutions.
In conclusion, this thesis identifies research gaps that warrant further exploration in the future
Deep Learning For The Classification of Lung Diseases Using Chest X-Rays
No full textThe discovery of X-rays marked a significant milestone in the field of medicine. One of the most common types of X-rays, the chest X-ray (CXR), allows doctors to examine an individual’s internal structure without surgery. Over the years, deep learning methods and algorithms have been developed to automate lung disease detection and identification. This paper introduces RADIA, a project that combines multiple deep learning techniques to identify abnormal areas and abnormal- ities in chest X-rays. RADIA builds upon previous studies conducted by the Stanford ML group, such as ChexNet and ChexPert. Our team utilized the ConvNeXt-Large, a deep learning convo- lutional model, implemented with a pre-trained ConvNext algorithm on the ImageNet database to classify various pathologies from public datasets like ChestX-ray14, CheXpert, MIMIC-CXR, PadChest, and VinDr-CXR, as well as a private dataset obtained from the Picture Archiving Com- munication System (PACS) at Verdun and Notre Dame Hospitals in Montreal in the collaboration with CIUSSS (Centre Integre Universitaire de Sante et de Services Sociaux du Centre-Sud-de-l’Ile- de-Montreal) and valuable consultants from the radiology team at Notre Dame Hospital contributed to the project’s success. Our team employed image enhancement and augmentation techniques to create various image versions. We used different and novel approaches to address the challenges, and the results were evaluated using metrics such as AUC, F1, and G-Means to analyze performance with imbalanced input data. It is essential to note that the project’s development extends beyond the creation of a web tool based on deep learning techniques. Our future plans involve building a decision helper that combines inference models and web tools to assist healthcare professionals
Advancing Shoreline Oil Spill Response with Eco-friendly Functional Nanomaterials
Get PDFMarine oil spills can cause serious environmental damage. When the spilled oil reaches the coast, it can result in negative implications for the coastal ecosystems. Coastal oil cleanup operations are often expensive and time-consuming. Surface washing can be used following an oil spill to enhance the removal of stranded oil from coastal surfaces. Surface washing agents (SWAs) are typically applied directly on stranded oils and oil is then flushed with ambient water to remove the oil and direct it to a controlled area for physical recovery. However, there is still a gap between the available SWAs and the increasing application need. Many SWAs have been produced for the treatment of oiled shorelines. Although the toxicity of some SWAs is moderate, they can still have a potentially adverse impact on the shoreline environment after application. Moreover, the effluents after washing can be further recovered through appropriate disposal to avoid secondary pollution.
This thesis presents the development of multiple environmentally friendly surface washing fluids that demonstrate commendable performance, reusability, and remarkable stimuli-responsiveness. The efficacy of these washing fluids was meticulously assessed under various environmental conditions. To gauge their impact on selected species, biotoxicity experiments and modeling were conducted. Molecular dynamic and thermodynamic modeling was utilized to unveil the intricate mechanism of oil removal. Additionally, post-treatment methods were explored to curtail potential secondary pollution stemming from washing sludge. Notably, the stimuli-responsive nature of these washing fluids played a pivotal role in generating clean supernatants with minimal turbidity and oil content. In a bid to harness water motion for both physical and chemical actions, a piezocatalytic washing fluid was also conceived. This innovative fluid demonstrated the capability to degrade oil into low molecular-weight hydrocarbons. Overall, this thesis holds immense value in significantly benefiting shoreline oil spill response by enriching cleanup techniques, reducing environmental impact, and enhancing cost-efficiency. Moreover, this thesis contributes to enhancing oil spill preparedness and response capacity and safeguarding valuable coastal regions
Adaptive Differential Privacy for Decentralized Mobility Data Sharing and Forecasting
Get PDFMobility data is the cornerstone of crucial applications, including traffic monitoring, crowdsourcing, and social networks. However, research shows that publishing accurate mobility data aggregate may jeopardize the participants’ privacy. As a robust and rigorous technique, differential
privacy provides a quantifiable protection guarantee by injecting enough noise into the aggregates to make them resilient to privacy attacks while allowing learning and analysis.
The application of differential privacy raises two challenges that stem from mobility data characteristics. First, mobility data is usually spread across multiple organizations, whereas standard differential privacy relies on a centralized trusted curator. Secondly, mobility data is typically sequential, while the guarantee provided by differential privacy degrades with consecutive aggregating of the sensitive data. This thesis tackles these challenges for two application scenarios: decentralized
mobility aggregate sharing and forecasting.
We leverage a distributed variant of differential privacy to enable decentralized mobility aggregate sharing where each organization obfuscates its dataset locally before sending it to the data curator. We use a sliding window approach to allocate the privacy budget to tackle the consecutive data access challenge. Moreover, we design an approximation strategy to calculate the closest private statistics to the current timestamp. We formally prove the privacy guarantee of our algorithms.
Finally, we demonstrate that our solution enables decentralized statistical release with a robust privacy guarantee on two datasets.
Before addressing the privacy aspect of distributed mobility forecasting, we design a mobility vertical federated forecasting (MVFF) framework that allows the learning process to be jointly conducted over vertically partitioned data belonging to multiple organizations. Since each organization only holds a location domain subset, none can tackle a forecasting model that covers the whole location domain. Moreover, distributed mobility data compromises the spatio-temporal correlation
between locations hindering learning. Hence, reducing the forecasting accuracy. MVFF uses a local learning model for each organization to extract the embedded spatio-temporal correlation between its locations. A global model synchronizes with the local models to incorporate the correlation between all the organizations’ locations. We investigate the performance of MVFF under four variations of local and global models. We compare the MVFF’s performance to two other federated frameworks on real-life datasets: New York Bike and Yelp reviews, achieving better performances.
Finally, we design two adaptive differential privacy budget algorithms for each organization participating in collaborative mobility forecasting. We define a new metric to assess the different organizations’ participation levels in the learning task and adjust the privacy budget accordingly.
Then, we adapt each organization’s privacy protection level (privacy budget) to the accuracy dynamics of the learning task. Lastly, we empirically evaluate our adaptive differential privacy budget algorithms using MVFF and two real-world datasets: a trajectory dataset collected in New York and Beijing over multiple months and a Yelp business review dataset
Video Atlantis, or (Post-)Soviet Small-Screen Cultures at the End of the Cold War
No full textArguing that video was part and parcel of the postsocialist condition marked by liberalization and disintegration of the so-called ‘Second World,’ this dissertation explores the history of analog video technologies, distribution, and consumption in the (ex-)Soviet Union in the 1980s and 1990s. (Post-)Soviet video, as this study emphasizes, represented a particular glocal formation that was a frontier of the Cold War economies and cultures since the 1960s, as well as their dismantling during the 1991 dissolution of the Soviet Union. First, while I trace Soviet video engineering projects, tele-video practices, and early consumer-grade video cultures, I outline the visions, both utopian and dystopian, that informed the meanings and usage of and policies around video as a technology. In particular, I foreground how video in the Soviet context emerges as a medium facilitating socialist and internationalist connectivity and offering a window into the world. Second, with an emphasis on how video reshaped the encounters between the (post-)Soviet and global realities, the study examines cross-border networks of video circulation and the rise of local video distribution infrastructure, including exhibition spaces, media bazaars, and television programming. Finally, this research brings translation as a cultural negotiation and actual language transfer to the fore, examining the crucial sites of (post-)Soviet video—children’s media and action cinema. With translation as a tool and object of the analysis, I interrogate the shifting geocultural dynamics of screen flows and politics of attributing cultural and aesthetic worth under (post-)socialism, as refracted through these media forms, video, and the reconfigured Cold War East-West divisions
Forecasting the Value-at-Risk of an Equity Portfolio: A Recurrent Mixture Density Network Approach
Get PDFThe value-at-risk is a useful metric employed by financial institutions to measure the risk of a portfolio. However, accurately forecasting the value-at-risk is difficult, as it requires predicting the returns of the portfolio's assets. Forecasting asset returns is particularly challenging due to their stochastic nature and the presence of 'stylized facts' such as heteroskedasticity, fat tails, and skewness in stock returns series. This thesis considers modeling the assets returns using a recurrent mixture density network, which has been previously proposed to model other financial time-series. In this thesis, we propose an improved recurrent mixture density network architecture, as well as a pretraining method to improve the numerical stability and convergence speed of the model. We also propose the Copula-S-RMDN-GARCH, which extends the current recurrent mixture density network architecture to multivariate settings. We compare the value-at-risk forecast obtained with the Copula-S-RMDN-GARCH with the forecasts obtained from a Copula-AR-GARCH
Graph Representation Learning for Classification and Anomaly Detection
Get PDFGraph-structured data is ubiquitous across diverse domains, including social networks, recommendation systems, brain networks, computational chemistry, biology, sensor networks, and transportation networks. Graph neural networks have recently emerged as a powerful paradigm for the analysis of graph-structured data due to their ability to effectively capture complex relationships and learn expressive graph node representations through iterative aggregation of information from neighboring nodes. These learned representations can then be used in various downstream tasks such as node classification and anomaly detection.
In this thesis, we introduce a graph representation learning model for semi-supervised node classification.
The proposed feature-preserving model addresses the challenges of oversmoothing and shrinking effects by introducing a nonlinear smoothness term into the feature diffusion mechanism
of graph convolutional networks. We conduct comprehensive experiments on diverse benchmark datasets demonstrating that our approach consistently outperforms or matches state-of-the-art baseline methods. Inspired by the concept of implicit fairing in geometry processing, we also propose a graph fairing convolutional network architecture for semi-supervised anomaly detection. The proposed model leverages a feature propagation rule derived directly from the Jacobi iterative method and incorporates skip connections between initial node features and each hidden layer, facilitating robust information propagation throughout the network. Our extensive experiments on five benchmark datasets showcase the superior performance of our graph fairing convolutional network
compared to existing anomaly detection methods. In addition, we propose an unsupervised anomaly detection approach on graph-structured data by designing a graph encoder-decoder architecture
and a locality-constrained pooling strategy. This pooling mechanism extracts local patterns and reduces the impact of irrelevant global graph information, enhancing the discriminative power of the learned features. In the decoding phase, an unpooling operation followed by a graph deconvolutional network reconstructs the graph data. Extensive experiments on six benchmark datasets demonstrate that our graph encoder-decoder model outperforms competitive baseline methods
A 5G Security Recommendation System Based on Multi-Modal Learning and Large Language Models
Get PDFDeploying 5G networks on top of cloud-native environments provides unique benefits including
cost-effectiveness, flexibility, and scalability. However, the increased complexity of a cloud-native
5G deployment also brings new security challenges to existing solutions for security monitoring
and security auditing. A security analyst may need to analyze events coming from various
sources, such as security monitoring (e.g., Falco) and security auditing (e.g., Kubescape) systems
deployed at both the 5G and cloud (container) levels. Understanding the relationships between
events coming from those different sources is usually challenging since those security solutions
may have very different monitoring/auditing criteria and scopes. Relying on manual analysis and
domain knowledge may also be too slow and error-prone considering the sheer scale of a cloud-native
5G deployment. In this paper, we propose 5GSecRec, a 5G security recommendation system
that leverages multi-modal learning to correlate alerts from four aspects of a cloud-native 5G deployment,
i.e., security monitoring and security auditing systems, deployed at both 5G/Kubernetes®
levels. Also, 5GSecRec further eases security analysts’ job by answering their questions expressed
in a natural language (e.g., “What is the impact of a Kubernetes alert on the 5G level?”) using
large language models (LLMs) fine-tuned with the learned knowledge about correlated alerts. We
implement 5GSecRec based on free5GC, Kubernetes, and LLMs from HuggingFace, and our experimental
results demonstrate the effectiveness of our solution (e.g., up to 89.5% of correlation
accuracy, and comparable question-answering performance to ChatGPT but without data confidentiality
and privacy concerns)