教学课程资源库 更多>>

关键词： Computer science  

关键词： Computer science   Medical imaging  

摘要： Multiple Sclerosis (MS) is a demyelinating disease of the central nervous system which affects nearly 1 million people in the US. Magnetic resonance imaging (MRI), as the most sensitive non-invasive technique, is used to visualize the disease course. In this dissertation, I present techniques to improve automated quantification of MS injuries based on MRI. Currently, focal white matter (WM) lesions and cortical gray matter (cGM) injuries are considered important biomarkers for diagnosing MS. Among these, WM lesion quantification has attracted more attention and there are many proposed methods focusing on WM lesion segmentation. However, due to the difficulty of this task, there is no fully-satisfactory method available to the community. To improve the performance of MS lesion segmentation, I develop a deep neural network and propose a data representation named 2.5D stacked slices. This approach achieved state-of-the-art performance for MS lesion segmentation and is currently listed at the top of one MS lesion segmentation challenge leaderboard. On the other hand, evaluation of cortical thinning, as one aspect of cGM injury quantification, can be improved by lesion inpainting. Towards this goal, I introduced the ``edge prior'' to guide the lesion inpainting process. With the edge prior, the deep networks are able to inpaint the images and better preserve the boundaries between tissues. Further, to make the proposed methods practical for clinical use, I explored the robustness of these deep learning models. One aspect is the models' robustness to adversarial attacks. Another aspect is the models' performance change on unseen domains (e.g., different clinic sites). To generalize the models to unseen datasets, I use many categories of data augmentations during the training process, assuming these augmentations can fill the gap between domains. Also, to address MRI sequence variations across sites, I apply modality dropout and let the models learn to make reasonable pred

关键词： Computer science   Mathematics   Artificial intelligence   Data models   Datasets   Hypotheses   Dissertations & theses   Data collection   Informatics   Visualization   Semantics   Bias  

摘要： Understanding highly-dimensional data sets is a complex task for many scientists, engineers, and intelligence analysts. Traditionally, this problem has been tackled with linear pipelines that rely on mathematical models and algorithms to summarize relationships and structure, producing a visual representation of the data in a collapsed, low-dimensional form. The main issue with these traditional pipelines is that they are driven solely by algorithms or models, and without a human in the loop, they can potentially limit sense-making by masking expected or known structure in the data. In recent years, Semantic Interaction has become a promising approach as a user interaction methodology for model steering in Visual Analytics systems, as it provides mechanisms with which to adjust the parameter space, explore data, and test hypotheses. Under the paradigm of Semantic Interaction, users can steer model parameters and explore data representations without leaving the visual space, thus combining algorithms and models with expert human judgment. Semantic Interaction systems need to invert the computation of one or more mathematical models to support a bidirectional structure within their pipelines to facilitate this interaction modality. For example, dimensionality reduction and clustering are frequently used to explore multidimensional data in Visual Analytic systems and are typically always present in Semantic Interaction systems. Since users interact with clustered data in its compressed form, the system needs to link this compressed form to the original high dimensional representation to affect the model and algorithms from within the visualization. The necessity of this reverse link from the low-dimensional representation to the high-dimensional input space requires that Semantic Interaction pipelines be bidirectional. Most examples of Semantic Interaction systems make use of simple and interpretable linear models for dimensionality reduction and clusterings such as LD

关键词： Computer science   Artificial intelligence  

摘要： Early attack identification is critical to secure the network exposed to various cyber-attack risks in modern times. Anomaly detection is also an important task, but devising a relevant strategy against the malicious event may be limited due to lacking attack information. For early attack identification, the main challenges are (1) imbalanced class data, (2) limited availability network datasets for early identification, and (3) determining variables to improve performance. To deal with these challenges, this research will explore early-identification of attacks with the following two angles: (1) Deep learning models including ANN, RNN; (2) Data augmentation for balanced data. Based on the analysis of the result of the experiments, this study will develop a deep learning model for early attack identification for validation purposes.

关键词： Mathematics   Theoretical mathematics   Computer science  

摘要： This dissertation is devoted to the development the augmented matched interface and boundary(AMIB) method and its applications for solving interface and boundary value problems. We start with a second order accurate AMIB introduced for solving two-dimensional (2D) elliptic interface problems with piecewise constant coefficients, which illustrates the theory of AMIB illustrated in details. AMIB method is different from its ancestor matched interface and boundary (MIB) method in employing fictitious values to restore the accuracy of central differences for interface and boundary value problems by approximating the corrected terms in corrected central differences with these fictitious values. Through the augmented system and Schur complement, the total computational cost of the AMIB is about $O(N \log N)$ for degree of freedom $N$ on a Cartesian grid in 2D when fast Fourier transform(FFT) based Poisson solver is used. The AMIB method achieves $O(N \log N)$ efficiency for solving interface and boundary value problems, which is a significant advance compared to the MIB method. Following the theory of AMIB in chapter 2, chapter 3 to chapter 6 cover the development of AMIB for a high order efficient algorithm in solving Poisson boundary value problems and a fourth order algorithm for elliptic interface problems as well as efficient algorithm for parabolic interface problems. The AMIB adopts a second order FFT-based fast Poisson solver in solving elliptic interface problems. However, high order FFT-based direct Poisson solver is not available in the literature, which imposes a grand challenge in designing a high order efficient algorithm for elliptic interface problems. The AMIB method investigates efficient algorithm of Poisson boundary value problem (BVP) on rectangular and cubic domains by converting Poisson BVP to an immersed boundary problem, based on which a high order FFT algorithm is proposed. This naturally allows for fulfilling a fourth order fast algorithm for so

关键词： Physics   Energy   Electromagnetics   Thermodynamics   Computer science   Computer engineering   Electrical engineering  

摘要： Spin-electronics or “spintronics” promises a new generation of low-power, high-speed, non-volatile memory and logic devices. Almost all existing and planned spintronic devices operate on controlling magnetization, preferably by electrical means alone. In this context, Cr2O3 (chromia) based heterostructures have emerged as a promising candidate. Chromia is magnetoelectric, which provides a direct coupling between an applied electric field and an induced magnetic moment. Moreover, chromia has an uncompensated moment at its interface, known as boundary magnetization. This magnetization is closely associated with the antiferromagnetic domain state of the chromia which can be electrically switched. The boundary magnetization can therefore be used as a logical bit in spintronic devices. While the equilibrium boundary magnetization of chromia is strictly tied to its domain state, non-equilibrium states may be created. For example, in the presence of an adjacent ferromagnet, exchange coupling at the interface may stabilize the interfacial domain state of the chromia against the rotation of the bulk of the crystal when it is subjected to critical magnetoelectrical switching fields. This results in an antiferromagnetic interface in a nonequilibrium state which is incommensurate with the underlying bulk. Upon magnetically cycling the ferromagnet, the antiferromagnet then relaxes, which in turn is reflected in the exchange bias of the ferromagnetic hysteresis loop. In other cases, the interfacial exchange locks the antiferromagnetic interface to rotate along with the ferromagnet. This causes a competition between the exchange energy at the interface and the anisotropy energy of the crystal. At the crossover point between these two mechanisms there occurs a simultaneous disappearance of exchange bias with a more than twofold increase in coercivity of the ferromagnetic hysteresis loop. Due to its relatively high Néel temperature, chromia has the advantage of potential operation a

关键词： Aircraft   Behavior   Software   Cameras   Metadata   Remote sensing   Artificial intelligence   Sensors   Boxes   Engineering   Algorithms   Clustering   Surveillance   Feasibility studies   Visualization   Virtual reality   Aerospace engineering   Computer science  

摘要： Remotely piloted aircraft systems (RPAS) have introduced a new ability to quickly deploy low-cost, fully or partially autonomous aerial sensor platforms which has created new intelligence, surveillance, and reconnaissance capabilities in various domains using cameras which are ubiquitous in most RPAS. Despite the utility of these aerial sensor systems, the full motion video (FMV) they acquire presents a big data challenge for operators as they generate large volumes of data that are impractical to analyze using current workflows due to excessive time requirements, computational resources, cost, or the availability of human analysts. Additionally, moving the camera rather than having a static network of stationary cameras, complicates the data processing steps required to generate valuable outputs. In order to address this big data challenge, various artificial intelligence (AI) based algorithms and data analytic workflows that can extract useful insights and knowledge from large amounts of complex and ambiguous FMV data streams from airborne sensors are developed and assessed. A data acquisition campaign was launched resulting in a dataset consisting of 33 flights recording approximately eight and a half hours of RPAS acquired FMV to assess the suite of AI-based algorithmic tools. Some of the tools useful for analyzing aerial FMV include object detection, and tracking namely to conduct pattern of life (POL) analysis for which aerial sensors mounted on RPAS are well suited for as they capture spatiotemporal information crucial to understanding the context of a scenario. Analysis and interpretation of the acquired dataset revealed that state of the art performance was achieved using the AI-based tools when the RPAS was deployed under an altitude of 30 m, at a velocity of under 7 m/s, and at pitch angles ranging from 25° to 65° while acquiring FMV at a resolution of 4.16 MP. The POL analysis conducted on two flights proved the two developed feature engineering based wo

关键词： Computer science   Remote sensing   Artificial intelligence  

摘要： Today, significant research efforts are spent on designing machine learning (ML) models to extract useful information from data. While ML models have shown superior performance in research studies, deploying them in practice is a challenging process. My dissertation considers the task of integrating ML models to real-world distributed systems, focusing on a number of sensing and monitoring applications. For such applications, ML deployments must overcome the dual challenge of heterogeneity and scalability. Heterogeneity refers to the need for ML models to be customized for instances of the application (e.g., a specific geographic location). Scalability refers to the challenges of training and deploying distinct models across a large number of nodes. In this dissertation, through specific case studies of high-impact problems, I demonstrate how I address these challenges using a system approach. My research considers both physical sensing and cyber monitoring applications. In physical sensing applications, heterogeneity mainly comes from sensors’ local environments, while in cyber monitoring scenarios, human factor is the key source of heterogeneity. In both types of applications, I study how heterogeneity affects the ML system design and evaluation, and propose new methods to address the dual challenges of heterogeneity and scalability. First, for the task of spectrum monitoring in cellular networks, I present a novel system design that can efficiently train and deploy deep neural network (DNN) models on a large number of spectrum sensors. Since each sensor observes complex, heterogeneous, and time-varying spectrum data, it is hard to train and deploy accurate DNN-based spectrum sensing models at scale. My work addresses this challenge by leveraging the hierarchical network structure of cellular networks, where I build context-agnostic models for spectrum usage and apply transfer learning to minimize training cost and dataset constraints. Then, I consider the problem

关键词： Neurosciences   Mathematics   Computer science   Medical imaging   Behavior   Accuracy   Principal components analysis   Datasets   Memory   Brain research   Flexibility   Methods   Correlation analysis   Algorithms   Clustering   Time series   Intelligence  

摘要： Recent advances in neuroimaging study, especially functional magnetic resonance imaging (fMRI), has become an important tool in understanding the human brain. Human cognitive functions can be mapped with the brain functional organization through the high-resolution fMRI scans. However, the high-dimensional data with the increasing number of scanning tasks and subjects pose a challenge to existing methods that wasn’t optimized for high-dimensional imaging data. In this thesis, I develop advanced data-driven methods to help utilize more available sources of information in order to reveal more robust brain-behavior relationship. In the first chapter, I provide an overview of the current related research in fMRI and my contributions to the field. In the second chapter, I propose two extensions to the connectome-based predictive modeling (CPM) method that is able to combine multiple connectomes when building predictive models. The two extensions are both able to generate higher prediction accuracy than using the single connectome or the average of multiple connectomes, suggesting the advantage of incorporating multiple sources of information in predictive modeling. In the third chapter, I improve CPM from the target behavioral measure’s perspective. I propose another two extensions for CPM that are able to combine multiple available behavioral measures into a composite measure for CPM to predict. The derived composite measures are shown to be predicted more accurately than any other single behavioral measure, suggesting a more robust brain behavior relationship. In the fourth chapter, I propose a nonlinear dimensionality reduction framework to embed fMRI data from multiple tasks into a low-dimensional space. This framework helps reveal the common brain state in the multiple available tasks while also help discover the differences among these tasks. The results also provide valuable insights into the various prediction performance based on connectomes from different tasks.

关键词： Computer science   Artificial intelligence  

摘要： Poisoning attacks have been described as a grave danger to neural networks, they tend to change the complete meaning of the model and also define how the model interprets the data provided to it. Certain inputs are retrained after which poisoned data can be induced into the model. As the popularity of self-driving cars are growing exponentially, we need to look into every major flaw in the system and have security measures in place for the safety of the passengers. There have been limited studies on progressive poisoning attacks especially when it comes to autonomous driving vehicles. In this work, we first generate the poisoned data and also propose a generative method that will increase the generation of the poisoned data. This method of poisoning helps us determine the extent of inaccuracy and unexpected side effects when the poisoned data is fed to the model.