教学课程资源库 更多>>

关键词： Computer Science   Electrical Engineering   machine learning   online learning   multi-armed bandits  

摘要： Online learning has been widely used in modern machine learning systems. In spite of the recent progress in online learning, many challenges remain unsolved when online learning is applied to the complex practical scenarios. The key challenges can be summarized from the following perspectives. First, curse-of-dimensionality of the action space: to learn the optimal action, the learning algorithm needs to explore all the actions to some extent. Thus, large-scale action space makes the learning algorithm converge slowly and incur large cost in the cold-start cases. Consider the online recommendation systems, the number of users or the number of items is a large number compared to the finite exploration budget. Second, non-stationary environments: real-world learning systems are highly dynamic, which is reflected in the fact that users' preferences change over time due to various internal or external factors, and item popularity vary due to fast emerging events. Failing to be adaptive to the changing environment may lead to sub-optimal decisions. Third, computational resource constraints: online learning algorithms need to be updated with the arriving data stream, which may require heavy computations for each arriving new data. The frequent updates cost a large amount of computation resources and prevent online learning tools from real-time applications. In this dissertation, we aim at developing efficient online learning solutions, and more specifically multi-armed bandit solutions, to conquer the aforementioned challenges. First, we study the correlations between actions in order to reduce the action space complexity of the online learning problem. We start with a special case of correlations, that playing one action also generates the side observations of other actions. Then we generalize the result to a general form of correlations over actions. Second, we consider multi-armed bandit in a non-stationary environment such that the learning algorithm can automatically

关键词： Computer science  

摘要： From mobile computing to the largest leadership-class supercomputers, many-core accelerators are relied upon to provide the computational capabilities necessary to make today's applications possible. Machine learning, image processing, n-body simulations, and a host of other applications are increasingly reliant on the computational capability provided by many-core accelerators to achieve the performance necessary to target real-world problems. For modern applications, especially those running on leadership-class supercomputers where the number of GPUs can outnumber traditional CPUs three to one, effective exploitation of many-core compute resources is a must to achieve high efficiency. Effectively exploiting the additional parallelism provided by many-core accelerators requires developers identify where accelerator parallelization would provide benefit and ensure efficient interaction between the CPU and accelerator. While these issues appear straightforward and well understood, we have found that significant untapped performance opportunities still exist even in well-studied, heavily optimized, real world applications created by experienced developers. By addressing hidden performance opportunities, we were able to reduce execution time by up to 87% for the applications we have tested. In this dissertation, we develop a new performance measurement and modeling technique called the feed-forward measurement model (FFM) that exposes previously hidden performance opportunities and delivers actionable feedback to developers on the potential benefit if the problem were corrected. We first explore a set of real-world applications to identify hidden performance opportunities common among them, developing techniques that tools can implement to detect their presence. FFM refines and expands these detection techniques by including a model that can estimate the performance benefit of fixing these problems. We have created a performance tool called Diogenes that implements FFM

关键词： Disability studies   Information technology   Computer science  

摘要： The problem addressed in this study was to define the aspects that software professionals need to comply with accessibility mandates when creating software. The study was a qualitative phenomenological study using lived experiences of the study participants to determine what is needed to rectify the situation or to figure out what the issue might be and how to rectify the issue. The study used the population from LinkedIn’s JAVA software developer group. The data was collected by conducting and recording purposeful interviews of the participants with the recording transcribed through a transcription service. Grounded theory was the data analysis used with the following themes emerging from the data collected: Software accessibility is taken into consideration for apps, websites, and government projects (508 compliance) and Software accessibility is not taken into account at the point of creation, it is not in business or customer requirements, uncertain of tools needed to ensure that software is compliant (508 compliance).

关键词： Design   Computer science  

摘要： Throughout the History of 3D media production, rigging characters for animation has been a time intensive and technically complicated bottleneck in the production pipeline. A goal of tools creators has constantly been to create better tools to help make this process more accessible and faster. Many tools that are created focus on the creation of a control rig because it is the most easily automatable and the most technically involved step. However, before that step can be completed, the user must have a skeleton set up for a character. With the rise of machine learning techniques like deep learning and an increase in GPU power, the automation of skeleton creation and association has become more attainable. Through our research many attempts at automating the character deformations have been found, but there seem to be far fewer in terms of better automatic joint placement. This thesis proposes a method for creating a deep 3D convolutional neural network (3D-CNN) which can accurately place joints within a bipedal character based on the volume of its mesh. It focusses on bipedal characters with a specific set of joints to simplify the data collection and training processes. While the processes described in this paper could be greatly improved with more time, this serves as a proof of concept that 3D-CNNs can handle this task well.

关键词： Mathematics   Computer science  

摘要： Phase retrieval is a common problem in signal processing, with applications in astronomy, x-ray imaging, electron microscopy, and coherent diffraction imaging (CDI). Many models and algorithms exist for phase retrieval, yet few are designed to handle noise without imposing additional assumptions like signal sparsity. One recent algorithm for noisy phase retrieval which requires no underlying assumptions is the Gauge Dual Descent (GDD) algorithm, which iteratively denoises the desired signal. The GDD algorithm involves a sequence of eigenvalue problems which can be computationally expensive for large-scale signal recovery, requiring many matrix-vector products. Additionally, signal denoising progress tends to stall in the GDD algorithm prior to satisfying optimality-based termination conditions. To address these challenges, this dissertation provides the following contributions. First, we establish empirical termination conditions for the GDD algorithm based on primal and dual variable relative errors which suggest signal denoising progress has stalled. Next, we develop an adaptive strategy for handling the sequence of eigenvalue problems in the GDD algorithm. These contributions lead to the Improved Gauge Dual Descent (IGDD) algorithm. Numerical examples demonstrate that the IGDD algorithm requires 50–80% fewer matrix-vector products than the GDD algorithm for a variety of problems with low-oversampling.

关键词： Computer science  

摘要： Being a spatial autocorrelation visualization tool in recent years, hotspot is often used in various fields, such as disease analysis, crime analysis, and weather conditions analysis and prediction in a certain area. Most of the research in hot spot analysis is in applying the concept to a variety of fields and to gain insights on the statistical significance prevalent in the clustering of data. Only a few of them discussed the efficiency and optimization of the algorithm. Commonly, these kinds of analyses would be based upon a huge dataset about space and time, and the conventional algorithm would take too much time to get the results. This paper mainly discusses whether the algorithm can be processed in parallel with MATLAB and how to further optimize the algorithm to shorten the calculation time and obtain accurate outcomes faster. I will use the toolbox ‘parpool’ in MATLAB on a multi-core node to parallelize the conventional algorithm, and then take advantage of the basic idea of the 'R-tree' to further optimize the parallel algorithm. In the end, the results are satisfactory, because the conventional serial algorithm can be parallelized in MATLAB, and the time consumption was saved about five times compared to the original algorithm. When the algorithm was further optimized, its time consumption is saved about ten times. This paper will be helpful in saving time when doing similar computations and analyses in the future.

关键词： Distributed System   Network Accelerator   Programmable Networks   Programmable Switch   SmartNICs   Computer science   Computer science and engineering  

摘要： The continuing increase of data center network bandwidth, coupled with a slower improvement in CPU performance, has challenged our conventional wisdom regarding data center networks: how to build distributed systems that can keep up with the network speeds and are high-performant and energy-efficient? The recent emergence of a programmable network fabric (PNF) suggests a potential solution. By offloading suitable computations to a PNF device (i.e., SmartNIC, reconfigurable switch, or network accelerator), one can reduce request serving latency, save end-host CPU cores, and enable efficient traffic control. In this dissertation, we present three frameworks for building PNF-enabled distributed systems: (1) IncBricks, an in-network caching fabric built with network accelerators and programmable switches; (2) iPipe, an actor-based framework for offloading distributed applications on SmartNICs; (3) E3, an energy-efficient microservice execution platform for SmartNIC-accelerated servers. This dissertation presents how to make efficient use of in-network heterogeneous computing re- sources by employing new programming abstractions, applying approximation techniques, co- designing with end-host software layers, and designing efficient control-/data-planes. Our prototyped systems using commodity PNF hardware not only show the feasibility of such an approach but also demonstrate that it is an indispensable technique for efficient data center computing.

关键词： Information science   Artificial intelligence   Computer science  

摘要： In the Big Data era, there is a resurgence of interest in using Datalog to express data analysis applications that require recursive computations. However, the use of non-monotonic aggregates in recursion raises difficult semantic issues. Recent theoretical advances like monotonic aggregation and Pre-Mapability (PreM) provide the formal semantics for the usage of aggregates in recursive Datalog rules enabling the expression of a wide spectrum of advanced analytical tasks, such as graph analysis, data mining, machine learning and stream processing. In this dissertation, we explore opportunities and issues created by these advances, including the expressiveness of Datalog in advanced applications and their optimization to achieve superior performance and scalability. Firstly, we find that Datalog serves as an efficient query language that simplifies the writing of machine learning applications and provides a unified environment for their development and deployment on multiple platforms. Following this route, we propose a declarative machine learning framework of tested effectiveness on top of Apache Spark. We present an in-depth theoretical analysis that shows how key ML algorithms can be expressed and efficiently implemented by recursive Datalog programs that use aggregates in recursion, whereby achieving both formal and efficient operational semantics. We also present the compilation and optimization techniques we developed to support the complex recursive queries required by ML applications in distributed share-nothing architectures. Next we share some theoretical results to show that programs computing any aggregates on sets of facts of predictable cardinality are equivalent to stratified programs where the pre-computation of cardinality of the set is followed by a stratum where recursive rules only use monotonic constructs. Finally, we investigate how to improve the parallelism of semi-naive evaluation of recursive Datalog programs on shared-memory multi-core mac

关键词： Information technology   Computer science  

摘要： Software Defined Networks have emerged and developed to become a prevalent industry used infrastructure. The mitigation around DDoS (Distributed Denial of Service) attacks in SDN has been a big topic since the network type has come into the scope of day to day operations. The crossfire attack is a link flooding Distributed Denial of Service attack that increases the amount of benign traffic from a massively distributed botnet to congest a network link, also known as the target link. This spike in traffic is used to deplete the network resources allocated for the target link. This attack normally does not contain any malicious payloads, which makes detecting and mitigating more difficult. This research was inspired by the low likelihood of the ability to detect and defend against a crossfire attack in software defined networks and is focused on detection and mitigation of these attacks [1]. The environment created in this experiment uses SDN Switches on a specific network topology in addition to the ryu controller. The random forest machine learning model was also utilized to dynamically analyze traffic and classify when an attack was beginning to occur. When the classifier alerts the controller that the threshold is being reached for a particular target link the controller will find and deny flows from the source which has generated the most traffic and is not regularly generating traffic at such a high rate. Denying flows closer to the source limits the impact that a flood can have on legitimate traffic in the topology. The threshold set for the target link in this experiment is 50 MB. The model is able to classify when an interface needs to be denied at a rate of 90.625 percent. Considering this is a base work, accuracy could be improved by having additional functionality applied to the proposed design.

关键词： Computer science  

摘要： With the arrival of the data explosion era, data modelling/prediction has become one of the most important research areas. Many datasets in our real-life are essentially data streams and functional data. This thesis investigates fuzzy system-based regression approaches for these two representative datasets. In line with the data types, this thesis is divided into two scenarios. The first part focuses on the data stream regression problem with input/output being real vectors. It is widely known that evolving fuzzy systems (EFSs) are effective approaches for solving data stream regression problems, in that EFSs are structurally self-organized, capable of updating structure/parameters in an online manner, and acting as the one-pass approaches without the requirement of storing historical data. However, current state-of-the-art studies indicate that the existing evolving structure/parameters approaches would impose a negative impact on the optimality of EFSs. To our best knowledge, research on proposing optimal EFSs was still rare. Furthermore, selecting predefined thresholds to control the structure/parameters evolution of EFSs is of importance, which has not been systematically investigated thus far. In this part, this thesis focuses on addressing the aforementioned two EFSs-related problems, which might provide implications for the research on the data stream. From an optimality viewpoint, EFS learning approaches for both Takagi-Sugeno and Mamdani fuzzy systems, that is, local error optimization approach (LEOA) and identifying evolving Mamdani fuzzy systems from the parameter optimization aspect (EMFSPO), are proposed. To automatically tuning the thresholds, two approaches, i.e. the self-evolving fuzzy system (SEFS), and an extended work based on SEFS, that is, EFS with self-learning/adaptive thresholds (EFS-SLAT), are proposed. Finally, through a wide range of benchmark examples, LEOA, EMFSPO, SEFS, and EFS-SLAT are shown to be capable of improving the accuracy comp