教学课程资源库 更多>>

关键词： Computer science   Neurosciences   Robotics  

关键词： Computer engineering   Artificial intelligence   Computer science  

摘要： The roads and highways are a valuable asset for the state department of transportation. It takes massive investment and a huge amount of time to maintain the road assets condition. Therefore, it becomes important to automate the maintenance process with minimum manual inspection. There has been significant research in the domain of classical computer vision techniques and machine learning methods concerning highway and road assets maintenance. However, the time consumed to assess and maintain and the amount of manual inspection involved is considerably large. Although adding automation to speed up the inspection process has been investigated by many studies, they overshadow the importance of scalable and light frameworks detecting the assets like storm drain and drop inlet in real-time. Thus, this thesis focuses on integrating a reliable and scalable AI Deep Learning framework customized for highway assets localization and detection of the assets in a road infrastructure environment mitigating the need for large and bulky model sizes. Furthermore, utilizing the advantage of the less computational cost and the lightweight framework architecture along with reasonably higher accuracy, it is possible to build an end-to-end framework that supports object localization and detection followed by the inference on the mobile edge embedded platforms. In a nutshell, this thesis presents a Deep Learning localization platform, customized to predict the position or the location of the asset items on the Highways such as drop inlets and storm drains. Moreover, it also provides results on the scalability of the localization task to the multi-object detection task with the help the state of the art EfficientDet-D0 model with a test accuracy of 73.4% mAP on the annotated test dataset and achieving validation accuracy of 51.67% mAP on the customized merged data of highway asset item drop inlet and 5 other COCO classes for object detection application. Additionally, various analyses bas

关键词： Computer science   Artificial intelligence  

摘要： Recent years have witnessed the great success of cloud computing that is powered by data center systems. As an essential building block for many of our cyber infrastructures, the monitoring, security, and management of data center traffic is of ultimate importance. However, the ever-increasing scale and speed of data center networks (DCNs) pose various challenges to data center traffic monitoring and management. The traditional monitoring approaches are neither flexible nor efficient to meet the security and performance needs of DCNs. With the flexibility and programmability of Software-Defined Networking (SDN) technology, this dissertation presents efficient solutions to address DCNs monitoring and management challenges. First, we investigate the feasibility and efficiency of building software-defined monitoring (SDM) functionalities into network edges. For this purpose, We design and implement four measurement schemes in Open vSwitch (OVS), by either integrating forwarding and measurement functions into a pipeline, or decoupling them into parallel operations. We further include another eBPF-based approach in our comparative study. Then, we quantitatively show the various trade-offs among network performance and system overhead that different schemes strike to balance, and demonstrate the feasibility of instrumenting OVS with monitoring capabilities. Second, we leverage the flexibility and efficiency offered by SDM to enhance the security of DCNs. Accordingly, we design BotSifter, an SDN-based scalable, accurate and runtime bot detection framework for DCNs. To improve the detection scalability, BotSifter utilizes centralized learning with distributed detection by distributing detection tasks across the network edges. Furthermore, it employs multiple novel mechanisms for parallel detection of Command and Control (C&C) channels and botnet activities, which greatly enhances the detection robustness. Third, the traffic surges in DCNs are common, due to the attacking (e

关键词： Biomedical engineering   Computer science  

摘要： Fatigue is one of the most prevalent phenomena in human beings, and yet its detection is highly subjective and poorly understood. The phenomenon of fatigue has a huge impact on performance, the ability to execute tasks safely and correctly, and the ability to retain or secure a job. Fatigue can be classified into two types: physical and cognitive fatigue. Physical fatigue may occur due to excessive physical exertion, while cognitive fatigue may occur due to excessive mental exertion. Historically, these two types of fatigue have been studied independently. However, in the real world, although these often occur at the same time, it has not always been easy to distinguish them or to understand which type of fatigue impacts the other. Therefore, it is important to study them together rather than separately in order to enable effective intervention and rehabilitation, regardless of the type of fatigue. This thesis presents a system designed to aid in human-centered studies that focus on the intervention of fatigue. It proposes a system that can detect both types of fatigue using the same setup. The system monitors the state of the body using physiological wearable sensors, which help us analyze the changes in the body due to fatigue. Using this multi-sensory approach, the system detects both physical and cognitive fatigue while collecting contextual data that help understand and predict occurrences of fatigue. The sensors employed vary: some are specific for detecting cognitive fatigue, while some are specific for detecting physical fatigue. Others can be used to detect both types. This creates a computational challenge in terms of sensor fusion and data processing. To address this, the system uses a Dynamic Data Driven Application System (DDDAS) paradigm and context-aware fusion techniques. These techniques allow the system to handle the different combinations of sensors and detect fatigue with higher accuracy. The main contributions of this thesis can be summarized in

关键词： Computer science   Artificial intelligence  

关键词： Operations research   Applied mathematics   Computer science  

摘要： Existing nonlinear optimization methods have proven reliable over the past few decades for a wide range of applications but have critically relied on accurate function and gradient evaluations. Modern nonlinear optimization problems arising from machine learning and scientific computing applications are increasingly complex and large scale, which make accurate evaluations of the function intractable, but may only require an approximate solution that is within some large tolerance or neighborhood of the true solution. One natural approach towards cheapening the iteration is by utilizing inaccurate function and gradient information by sampling or allowing for computational noise. In this thesis, we develop scalable second-order methods for unconstrained optimization that can tolerate noise in the function and gradient. We consider the settings where the derivatives are available or unavailable (derivative-free optimization) under two frameworks of noise: (1) stochastic optimization, where the objective function is defined as the expected value of some function, where one can only obtain the sampled function and gradient; and (2) noisy optimization, where one obtains inaccurate function and gradient evaluations that are contaminated with computational noise. In the first project, we develop a new version of the L-BFGS algorithm for stochastic optimization. The standard L-BFGS method relies on gradient approximations that are not dominated by noise, so that search directions are descent directions, the line search is reliable, and quasi-Newton updating yields useful quadratic models of the objective function. All of this appears to call for a full batch approach, but since small batch sizes give rise to faster algorithms with better generalization properties, L-BFGS is currently not considered an algorithm of choice for large-scale machine learning applications. One need not, however, choose between the two extremes represented by the full batch or highly stochastic reg

关键词： Computer science   Applied mathematics   Artificial intelligence  

关键词： Physics   Condensed matter physics   Computer science  

摘要： The importance of topology in condensed matter physics has expanded in recent years as it has become a key piece of understanding the physics in various phenomena. In this thesis, we examine several examples of the power of topological concepts in defining and understanding certain systems and effects. We consider magnetic skyrmions, whirls of magnetic texture, the topology of which is signified by the topological charge, a quantity dependent on the configuration of the magnetic moments in real space. The stability granted by this topology makes skyrmions of interest for computing applications. For elongated skyrmions coupled to a superconductor, the existence of an energy gap can indicate that the region around the skyrmion is in a topological phase. This makes it possible for Majorana bound states to exist at the edges of the skyrmion. Majorana bound states provide great promise for quantum computing, as they behave as non-abelian anyons. We also study Berry curvature, for both electrons and magnons in momentum space, which leads to various versions of the Hall effect in both cases. These Hall effects, which can give rise to transverse currents, including charge and spin currents, could also be used for devices. In the electron case, we show how in a system of graphene atop a chromia substrate, the Chern number represents the topological nature of the system and signifies which effects will be present, such as the quantum anomalous Hall effect and valley-polarized quantum anomalous Hall effect. For an insulating noncollinear antiferromagnet, we calculate the magnon spin Nernst response for potassium iron jarosite KFe3(OH)6(SO4)2.

关键词： Mathematics   Computer science   Applied mathematics   Artificial intelligence  

摘要： Deep neural networks have emerged, in recent years, as incredibly powerful models in machine learning. Despite this, theoretical understanding of neural networks are lacking, or unsatisfying. In this thesis, we consider two problems in the theory of neural networks, focusing on function approximation via feed-forward neural networks. The first part of this thesis deals with understanding the approximation capacities of neural networks in terms of their depth. We first discuss this problem in the low-dimensional case, focusing on a specific class of functions, namely solutions to transport problems. We then move to the high-dimensional setting, a regime more relevant to machine learning. We look at neural networks in the frequency domain, and we offer explanations on why and when depth is essential to computational efficiency. In the second part of the thesis we look at the problem of optimization, and we discuss the optimization landscape for shallow neural networks, focusing on distribution-free results on existence of spurious minima regions. We finish with some concluding remarks and discussing a few open questions.

关键词： Computer science   Bioinformatics   Statistical physics   Biomedical engineering   Applied mathematics  

摘要： Motion analysis and studying the motion patterns of objects play a crucial role in a variety of applications. Object tracking, as a means for motion analysis, is an integral part of new technologies like aircraft and satellite navigation or biomedical applications. It entails locating objects in a video, registering them over time, and extracting their trajectories for downstream analysis. In biomedical applications, the target objects could be cells, sub-cellular organelles, pathogens, or multi-cellular organisms. Some examples of Subsequent downstream analysis in the biomedical application could be classifying healthy cellular behavior from infected ones, differentiating motion of an organism with and without the presence of certain drugs for drug effectiveness studies, or identifying dynamic changes in cellular morphology as a result of the pathogenic invasion. The first and most important step in this process is the segmentation and tracking of relevant objects or the identification of regions in images corresponding to relevant biological phenomena and tracking those phenomena over time. Current tracking and segmentation approaches in biological applications have numerous drawbacks that severely limit their utility. First, they often make strong \textit{a priori} assumptions about the data, limiting the efficacy of the approach to a narrow subset of biological/biomedical applications contexts specifically. Second, most state-of-the-art methods require a large amount of “ground truth”, manually annotated image data, to accurately train their models. This is a significant obstacle for tracking objects in different settings. Third, most modern approaches are designed to track no more than single objects, while in biological/biomedical applications, we need to analyze the motion dynamics of hundreds of objects simultaneously, with possible occlusions and deformations, to study the change in motion behavior. Here, we propose a scalable unsupervised biomedical motion