教学课程资源库 更多>>

关键词： Software   Neural networks   Support vector machines   Artificial intelligence   Computer science  

摘要： Machine learning, which has its origins at the intersection of computer science and statistics, is now a rapidly growing area of research that is being integrated into almost every discipline in science and business such as economics, marketing and information retrieval. As a consequence of this integration, it is necessary to understand how machine learning interacts with these disciplines and to understand fundamental questions that arise at the resulting interfaces. The goal of my thesis research is to study these interdisciplinary questions at the interface of machine learning and other disciplines including mechanism design/information elicitation, preference/choice modeling, and theoretical computer science.

关键词： Computer science   Information technology   Artificial intelligence  

摘要： This thesis is an investigation into the performance ramifications of making specialized component reservations for machine learning workloads in virtualized environments. The reliance of machine learning on floating-point operations makes graphics processing units an important part of processing these workloads quickly. While virtualization is one of the most widely-used consolidation techniques used in data centers of all sizes, compatibility issues between graphics processing units and virtualization have slowed the adoption of virtualization for machine learning workloads. To that end, this paper discusses the motivations and history behind virtualization and the application-specific acceleration devices used and how they are applied to machine learning on various public and private computing platforms. This is followed by a presentation of an experimental framework for testing the impact of controlling for non-uniform memory access when running machine learning workloads. Using this framework, a series of experiments were performed and documented in this thesis that test multiple placement configurations for graphics processing units in a virtualized system and how throughput of data from the host system to the device was affected. Current virtualization platforms offer recommendations to use these settings, but do not talk about the specific impacts of implementing them. Based on the results of the experiments, configuration parameters and placement recommendations are presented along with information about how these settings can help optimize the machine learning pipeline and the potential pitfalls to their use.

关键词： Chemistry   Computer science   Thermodynamics   Computational chemistry  

关键词： Health care management   Electronic health records   Data integrity   Medical research   Internet crime   Medical equipment   Malware   Privacy   Computer science  

关键词： Energy   Computer science   Electrical engineering  

摘要： The electricity is generated in bulk power plants and transported to the end-user through the transmission and distribution networks. The process incurs heavy losses adding to the operational costs. Secondly, fossil fuels dominate energy generation and are a major source of greenhouse gases. Thirdly, the power grid is vulnerable to natural disasters. The smart grid addresses these challenges by integrating distributed energy resources (DERs) in the distribution system closer to the load and with greater penetration of renewable energy. Renewable energy is key to cutting carbon emissions due to fossil fuel-based electricity generation and reducing operating costs. It can also effectively help speedier localized service restoration before the main supply becomes available after a disaster. However, renewable energy sources, such as solar and wind, bring in a new challenge of intermittent generation which can destabilize the power grid. As a result, to incentivize the user to do demand-side management, the power grid employs dynamic electricity pricing. In this context, this dissertation focuses on optimizing the demand and distribution of power in a smart grid. In the first case, we investigate profit maximization in geo-distributed data centers by optimizing power consumption through request routing and optimal resource allocation. Next, we extend this work by proposing a cooperative strategy where data centers incorporate grid stability into the optimization process. It enables the data centers to utilize lower power tariffs while minimizing geographic load balancing overheads. It also facilitates the grid to increase renewable penetration without catering to additional reserves. In the second case, this dissertation investigates the use of DERs to mitigate disaster effects through energy donation. Our approach simultaneously prioritizes critical load and provides relief to those in dire need while complying with resources and network constraints. Finally, this diss

关键词： Modulation   Feature extraction   Training   Image recognition   Convolution   Sun   Computer science   Indoor scene recognition   privileged information   attention transfer  

摘要： RGB-D scene recognition has achieved promising performance because depth could provide complementary geometric information to RGB images. However, the inaccessibility of depth sensors severely limits RGB-D applications. In this paper, we focus on depth privileged setting, in which depth information is only available during training but not available during testing. Considering that the information obtained from RGB and depth images are complementary while attention is informative and transferable, our idea is using RGB input to hallucinate depth attention. We build our model upon modulated deformable convolutional layer and hallucinate dual attention: post-hoc importance weight and trainable spatial transformation. Specifically, we use modulation (resp., offset) learned from RGB to mimic Grad-CAM (resp., offset) learned from depth, to combine the strength of dual attention. We also design a weighted loss to avoid negative transfer according to the quality of depth attention. Extensive experiments on two benchmarks, i.e., SUN RGB-D and NYUDv2, demonstrate that our method outperforms the state-of-the-art methods for depth privileged scene recognition.

关键词： Computer science   Artificial intelligence   Computer engineering  

摘要： While memory is fundamental in enabling intelligence, the development of neural memory architectures has largely fallen behind compared to the recent flourish of time-independent neural models. In this thesis, we contribute to the advancement of this field by proposing a novel neural memory model, Multigrid Neural Memory, in which we study in detail three orthogonal dimensions in building the system: architectural design, domain-agnostic processing, and local operators. First, we introduce a radical new approach to endowing neural networks with access to long-term and large-scale memory. Architecting networks with internal multigrid structure and connectivity, while distributing memory cells alongside computation throughout this topology, we observe that coherent memory subsystems emerge as a result of training. Our design both drastically differs from and is far simpler than prior efforts, such as the recently proposed Differentiable Neural Computer (DNC), which uses intricately crafted controllers to connect neural networks to external memory banks. Our hierarchical spatial organization, parameterized convolutionally, permits efficient instantiation of large-capacity memories. Our multigrid topology provides short internal routing pathways, allowing convolutional networks to efficiently approximate the behavior of fully connected networks. Such networks have an implicit capacity for internal attention; augmented with memory, they learn to read and write specific memory locations in a dynamic data-dependent manner. We demonstrate these capabilities on synthetic exploration and mapping tasks, where our network is able to self-organize and retain long-term memory for trajectories of thousands of time steps, outperforming the DNC. On tasks without any notion of spatial geometry: sorting, associative recall, and question answering, our design functions as a truly generic memory and yields excellent results. Second, we introduce a novel processing scheme that helps enab

关键词： Biostatistics   Standard deviation   Malaria   Simulation   Population   Vaccines   Computer science   Signal to noise ratio   Parasites   Estimates   Disease   Diagnostic tests   Microscopy   Methods   Genes   Bias  

摘要： This thesis is a compilation of three separate projects. Although the three projects are quite distinct in that they do not directly build on one another, the common thread is statistics and computation applied to medicine and biology. In the first project, we develop a Bayesian statistical model to quantify the size of and overlap between two populations from subsamples from each population. Our model produces joint estimates of the size of each population and the amount of overlap between the populations. We apply our model to compare the genetic repertoires of malaria parasites. In particular, we use our method to construct parasite genetic similarity networks with which we analyze the evolution and transmission of malaria in Ecuador before and after an outbreak in 2012-2013. We also use our method's inferences to argue that the polymerase chain reaction primers used to study an important malaria parasite gene are biased. In the second project, motivated by the COVID-19 pandemic, we develop a Bayesian statistical model to estimate vaccine effectiveness using data from a common vaccine effectiveness study design called a test-negative design. Since disease positivity is typically assessed using diagnostic tests that have imperfect sensitivity and specificity, our model adjusts for imperfect diagnostic tests along with other potential confounders. In addition, our framework allows data streams from multiple diagnostic test types to be rigorously combined which makes our model flexible for analyses and meta-analyses of test-negative designs. In the third project, we study image resolution metrics that are applicable to modern super-resolution fluorescent microscopy techniques. After showing that the primary existing quantitative image resolution metric called Fourier ring correlation is not ideal, we argue that image resolution in super-resolution microscopy might best be analyzed using information theory. Towards that end, we propose a mutual information-based imag

关键词： Tomography   Diabetic retinopathy   Retina   Image retrieval   Neural networks   Signal processing   Edema   Algorithms   Melanoma   Computer science   Computational physics   Applied physics   Medical imaging  

摘要： The computational approach to image acquisition and analysis plays an important role in medical imaging and optical coherence tomography (OCT). This thesis is dedicated to the development and evaluation of algorithmic solutions for better image acquisition and analysis with a focus on OCT retinal imaging. For image acquisition, we first developed, implemented, and systematically evaluated a compressive sensing approach for image/signal acquisition for single-pixel camera architectures and an OCT system. Our evaluation outcome provides a detailed insight into implementing compressive data acquisition of those imaging systems. We further proposed a convolutional neural network model, LSHR-Net, as the first deep-learning imaging solution for the single-pixel camera. This method can achieve better accuracy, hardware-efficient image acquisition and reconstruction than the conventional compressive sensing algorithm. Three image analysis methods were proposed to achieve retinal OCT image analysis with high accuracy and robustness. We first proposed a framework for healthy retinal layer segmentation. Our framework consists of several image processing algorithms specifically aimed at segmenting a total of 12 thin retinal cell layers, outperforming other segmentation methods. Furthermore, we proposed two deep-learning-based models to segment retinal oedema lesions in OCT images, with particular attention on processing small-scale datasets. The first model leverages transfer learning to implement oedema segmentation and achieves better accuracy than comparable methods. Based on the meta-learning concept, a second model was designed to be a solution for general medical image segmentation. The results of this work indicate that our model can be applied to retinal OCT images and other small-scale medical image data, such as skin cancer, demonstrated in this thesis.