教学课程资源库 更多>>

关键词： Applied physics   Theoretical mathematics   Computer science  

摘要： Random matrix theory has a long history. It was first introduced in mathematical statistics by John Wishart in 1928, and it gained attention during the 1950s due to work by Eugene Wigner studying the distribution of nuclear energy levels. A large number of physicists and mathematicians have been fascinated by random matrix theory, and after decades of study, it has matured into a field with applications in many branches of physics and mathematics. Nowadays, the subject is still very much alive with new and exciting research. Much of my PhD work has revolved around the study of random matrix theory. This dissertation gives a tour of my work on asymptotic theory of large random matrices and its applications in statistics, probability, and the theory of orthogonal polynomials, respectively.

关键词： Electrical engineering   Computer science  

摘要： Traditionally, datacenter networks are constructed with electronic packet switches, which forward data in a distributed manner. However, packet switches are running in to serious scaling challenges due to the massive growth in bandwidth requirements within major datacenters. One possible solution to this situation is to use optical circuit switching, which has significant potential for both power and cost savings. However, circuit switching is a completely different paradigm from packet switching, and as such effectively utilizing circuit switching requires work at all layers in the network, including the network architecture, switches, transceivers, network interfaces, and protocols. This dissertation validates the thesis that developing a NIC capable of precision admission control and characterizing its performance can lead to practical sub-microsecond circuit-switched networks at scale. Specifically, this dissertation presents novel contributions to two key system-level issues inherent in utilizing optical circuit switching in datacenter networking applications. The first contribution of this dissertation quantifies the system-level reconfiguration time of an optically circuit-switched link. This work is critical for understanding how to build optical links using high speed optical switches and how to integrate these links into a datacenter environment. This includes a discussion of system-level reconfiguration time and link-level measurements of an optically switched link, including bit error rate (BER) characterization of a 25 Gbps link utilizing a burst-mode receiver, switched by a nanosecond-scale silicon photonic switch. The second contribution of this dissertation is the development of a network interface controller (NIC) that can precisely control the injection of packets into an optically-switched network. This work describes and quantifies the performance of a novel high-performance, open-source, FPGA-based NIC called Corundum. This NIC is designed to pr

关键词： Computer science  

摘要： Deep learning widely applies to solve various problems in healthcare, robotics, and computer vision. Presently, an emerging deep learning application called "deepfake" has raised concerns about the multiple types of security threats that may pose severe harm to personal privacy and public safety. Deep convolutional neural networks like VGGNet and InceptionNet have recently set a proposal for detecting deepfake. The main challenge of these CNN-based algorithms is that they require extensive training datasets and high-end GPU resources. Furthermore, these studies mainly focus on identifying patterns in facial expressions in deepfake, and there are only very few studies on detecting audio fakeness. In this thesis, we propose a novel method for uni-modal or multi-modal deepfake detection with minimum resources. The proposed solution was designed with a Siamese network-based deepfake model with invariant of constructive loss and triplet loss. Contrastive loss uses the trained network's output for a positive example and calculates its distance to an instance of the same class and contrasts it with the range to negative samples. The triplet loss was computed by positioning the baseline that minimizes the distance to positive samples but maximizes the distance to negative samples. To test and validate our proposed model, we report our metrics like similarity score, loss, and accuracy on large-scale DFDC, Faceforensic++, and CelebDF datasets. We compared our method with state-of-the-art algorithms and confirmed that our overall accuracy is improved by 2-3% for deepfake detection.

关键词： Computer science  

摘要： This thesis focuses on the modelling, analysis and design of novel communication strategies for wireless machine-type communication (MTC) systems to realize the notion of Internet of things (IoT). We consider sensor based MTC devices which acquire physical information from the environment and transmit it to a base station (BS) while satisfying application specific quality-of-service (QoS) requirements. Due to the wireless and unattended operation, these MTC devices are mostly battery-operated and are severely energy-constrained. In addition, MTC systems require low-latency, perpetual operation, massive-access, etc. Motivated by these critical requirements, this thesis proposes optimal data communication policies for four different network scenarios. In the first two scenarios, each MTC device transmits data on a dedicated orthogonal channel and either (i) possess an initially fully charged battery of finite capacity, or (ii) possess the ability to harvest energy and store it in a battery of finite capacity. In the other two scenarios, all MTC devices share a single channel and are either (iii) allocated individual non-overlapping transmission times, or (iv) randomly transmit data on predefined time slots. The proposed novel techniques and insights gained from this thesis aim to better utilize the limited energy resources of machine-type devices in order to effectively serve the future wireless networks. Firstly, we consider a sensor based MTC device communicates with a BS, and devise optimal data compression and transmission policies with an objective to prolong the device-lifetime. We formulate joint optimization problems aiming to maximize the device-lifetime whilst satisfying the delay and bit-error-rate constraints. Our results show significant improvement in device-lifetime. Importantly, the gain is most profound in the low latency regime. Secondly, we consider a sensor based MTC device that is served by a hybrid BS which wirelessly transfers power to the devic

关键词： Mathematics   Operations research   Computer science  

摘要： Non-convex time-consuming objectives are often optimized using “black-box” optimization. These approaches assume very little about the objective. While broadly applicable, these approaches typically require more evaluations than methods exploiting more problem structure. In particular, often, we can acquire information about the objective function in ways other than direct evaluation, which is less time-consuming than evaluating the objective directly. This allows us to develop novel Bayesian optimization algorithms that outperform methods that rely only objective function evaluations. In this thesis, we consider three problems: optimization of sum and integrals of expensive-to-evaluate integrands; optimizing hyperparameters for iteratively trained supervised learning machine learning algorithms; and optimizing non-convex functions with a new efficient multistart stochastic gradient descent algorithm.

关键词： Biology   Bioinformatics   Computer science  

摘要： Throughout a human’s lifespan hematopoietic stem cells (HSCs) acquire somatic mutations. Some mutations can provide a proliferative advantage that eventually results in a substantial proportion of blood cells derived from a single mutant lineage. When these types of mutations occur in myeloid cancer-associated genes, it is referred to as clonal hematopoiesis (CH). The gene encoding DNA methyltransferase 3A (DNMT3A) is the most prevalently mutated gene in CH, but it is still poorly understood how these mutations offer a fitness advantage early on. This work explores how DNMT3A mutations drive the epigenetic dysregulation that results in a clonal expansion and shapes early hematopoiesis in humans. To investigate this phenomenon, an integrative single cell multi-omics approach was used to analyze a cohort of three individuals harboring DNMT3A mutations at the mutational hotspot R882C. Targeted genotyping of DNMT3A transcript allowed for identification and comparison of mutant and wild-type cells from the same individual. In addition, whole transcriptome and methylome status was captured to reveal a transcriptional priming bias and global hypomethylation in at least one patient. Transcription start sites (TSSs) of genes involved in differentiation, including HOX-related genes, were found to be differentially methylated between mutant and wild-type cells. Additionally, regions containing binding sites of transcription factors (TFs) involved cell fate [e.g. ZBTB1 and SMAD5] were found to be strongly hypomethylated in mutant cells. Two of the patients exhibited hypomethylation at reported association sites of polycomb group (PcG) proteins, EZH2 and SUZ12. Thus, altered DNA methylation (DNAme) of PcG protein association sites could provide a mechanism for transcriptional disruption in mutants. Importantly, an erythroid transcriptional priming bias was detected in mutant HSCs, providing direct evidence of transcriptional changes associated with DNMT3A mutations. This finding

关键词： Computer science  

摘要： Sparse matrix-vector multiplication (SpMV) is an essential computation in linear algebra. There is a well-known trade-off between operating on a dense or a sparse structure when performing SpMV. In the dense version of SpMV, useless operations are performed but the computation is amenable SIMD vectorization. In the sparse version, only useful operations are executed. However, an indirection array must be used, thus hindering the compiler's ability to perform optimizations that exploit the vector units available on the majority of modern processors. Our process automatically builds sets of regular sub-computations from the irregular sparse data structure. We mine for regular regions in the irregular data structure, grouping together non-contiguous points from the reorderable set of coordinates representing the sparse structure. The coordinates become partitioned into groupings of coordinates of pre-defined shapes using polyhedra. This partition models the exact same points from the input set of coordinates in a way that is specialized to the input's sparsity pattern. Once we have obtained a partition of the points into sets of polyhedra, we then scan these polyhedra to synthesize code that does not store any coordinates of zero-valued elements and does not require any indirection array to access data, thus making it amenable to SIMD vectorization.

关键词： Computer science  

摘要： Software engineering is a discipline closely involved with human activities. How to help software developers produce better software more efficiently is the core topic of software engineering. Prioritizing tasks can efficiently reduce the human efforts and time required to retrieve certain relevant software artifacts (e.g. relevant research papers, security vulnerabilities in the code, test cases detecting faults, etc.). Many of such prioritization problems in software engineering can be generalized as the total recall problem. The total recall problem aims to optimize the cost for achieving very high recall—as close as practicable to 100%—with a human assessor in the loop. This problem has been explored in information retrieval for years, and the state of the art solution with active learning and natural language processing aims to resolve the following challenges: Higher recall and lower cost. Among a finite number of tasks, which task to be executed first so that most of the relevant artifacts can be retrieved earlier? When to stop. At what point, there is no need to execute the remaining tasks? Human error correction. Are all the tasks executed correctly? How to identify wrongly executed tasks and correct them? Although being explored for years, all three challenges have much room to be improved. We claim that a broad class of software engineering problems, where humans must apply their expert knowledge to identify target software artifacts from a tediously large space, can be generalized as the total recall problem and be better solved by active learning. To support this claim, we studied the existing state of the art techniques in solving the total recall problems, and proposed a new active learning-based framework, FASTREAD, to resolve the three challenges of total recall. FASTREAD was tested on software engineering literature review datasets and were found to outperform the existing state of the art solutions. We also conducted our own literature review stud

关键词： 功能对等理论   英汉翻译   计算机科学技术文本   翻译实践报告  

摘要： 计算机科学领域的各项技术,尤其是最近几年出现在人们生活中的人工智能技术,以超乎想象的速度推动着社会向前发展。人工智能是战略制高点,也是未来互联网发展的技术核心,并将引发产业结构的深刻变革。尤金·奈达的功能对等理论的核心原则就是使译文读者获得与原文读者相同的反应。为了使科技英语翻译既能准确传达科技文本中的信息,又能使原文和译文实现最自然、最切近的对等,本文在科技英语文本的翻译中尝试运用了尤金·奈达的功能对等理论。该实践报告主要分析了科技文本的翻译过程。本文首先介绍了翻译材料的来源以及语言特点,如在词语层面呈现多专业术语及多名词化结构,在语句层面呈现多被动句及多复杂句,在语篇层面逻辑严谨等;第二说明了翻译的过程,包括译前准备、翻译过程中工具的选择与翻译计划的制定以及译文的质量控制;第三根据功能对等理论,从词语、语句及语篇三个层面对计算机科技文本的翻译运用不同的方法进行分析研究;最后在报告的结论部分中总结了整个翻译过程以及研究的发现和不足。本文的翻译实践表明,功能对等理论对科技英语翻译质量的提升有积极的指导作用,同时本文为计算机科学技术文本的翻译实践提供参考并且激发对计算机技术相关文本的翻译研究。

关键词： Computer science  

摘要： For decades, developers have been productive writing software by composing optimized libraries and functions written by other developers. Though hardware trends have evolved significantly over this time---with the ending of Moore's law, the increasing ubiquity of parallelism, and the emergence of new accelerators---many of the common interfaces for composing software have nevertheless remained unchanged since their original design. This lack of evolution is causing serious performance consequences in modern applications. For example, the growing gap between memory and processing speeds means that applications that compose even hand-tuned libraries can spend more time transferring data through main memory between individual function calls than they do performing computations. This problem is even worse for applications that interface with new hardware accelerators such as GPUs. Though application writers can circumvent these bottlenecks manually, these optimizations come at the expense of programmability. In short, the interfaces for composing even optimized software modules are no longer sufficient to best use the resources of modern hardware. This dissertation proposes designing new interfaces for efficient software composition on modern hardware by leveraging algebraic properties intrinsic to software APIs to unlock new optimizations. We demonstrate this idea with three new composition interfaces. The first interface, Weld, uses a functional intermediate representation (IR) to capture the parallel structure of data analytics workloads underneath existing APIs, and enables powerful data movement optimizations over this IR to optimize applications end-to-end. The second, called split annotations (SAs), also focuses on data movement optimization and parallelization, but uses annotations on top of existing functions to define an algebra for specifying how data passed between functions can be partitioned and recombined to enable cross-function pipelining. The third, ca