教学课程资源库 更多>>

关键词： Computer science   Artificial intelligence   Transportation  

摘要： In this thesis, we enabled automatic sensing for intelligent vehicles that utilize their perception ability through sensors to predict correct driving behaviors for safe driving. First, we focused on developing deep learning models that learn from real world driving data collected by driving on local roads and on highways. The models are trained, tested, and applied to predict the steering angle of vehicles. Specifically, we leveraged convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to develop deep learning models that make use of real driving streaming data and respective steering angles to learn. We also studied the literature to build motion prediction and path planning modules that can be deployed to assist driving and promote safe driving. Preliminary experiments were conducted and experimental data is collected to examine the performance of the selected motion prediction and planning algorithms. We then discussed our future work at the end of this thesis.

关键词： Geophysics   Geology   Optics   Petroleum engineering   Computer science   Acoustics  

摘要： Distributed acoustic sensing (DAS) is a method that utilizes the Rayleigh backscattering of laser light pulses generated at numerous randomly spaced scattering points along a fiber optics cable, and then recorded by an “interrogator box”, where it’s measured by a computer known as an Optical Time Domain Reflectometer. Using the principles of two-way travel time, we can locate where backscatter originates along the fiber and any subsequent changes in arrival time, phase, or frequency of input laser light, which act as a measurement of the strain rate acting on the fiber cable. DAS technology can be useful in petroleum industry applications such as vertical seismic profiling (VSP), microseismic detection and imaging, and hydraulic fracture monitoring. In this study, we analyze a unique DAS dataset recorded from a producing horizontal oil well with an open-hole completion design, meaning the lateral section is not hydraulically stimulated. The DAS fiber rod is coupled directly to the reservoir under its own weight, directly on top of the reservoir rock. The well’s production alternated on and off multiple times over a 60-hour window, while DAS data recorded continuously. Signal analysis was performed on the data to identify and differentiate signal events from noise, and determine an optimal workflow for imaging the data, suitable for the large volumes (e.g., 7 Terabytes) of data that DAS creates. We have shown it is possible to use an extremely large dataset reduced by more than 99%, while retaining clear acoustic signal in the unenhanced data. This study positively identifies and interprets fluid-flow signal events visible on the DAS data during both phases of production and models the interpreted DAS signal to physical phenomena occurring in the wellbore. Understanding the acoustic fluid-flow signals in DAS data in a passive, time-lapse environment has potential to impact oil field and reservoir development including optimizing production, fluid injection, and knowl

关键词： Computer engineering   Computer science   Artificial intelligence  

摘要： Sparse matrices, more specifically Sparse Matrix-Matrix Multiply (SpGEMM) kernels, are commonly found in a wide range of applications, spanning graph-based path-finding to machine learning algorithms (e.g., neural networks). A particular challenge in implementing SpGEMM kernels has been the pressure placed on DRAM memory. One approach to tackle this problem is to use an inner product method for the SpGEMM kernel implementation. While the inner product produces fewer intermediate results, it can end up saturating the memory bandwidth, given the high number of redundant fetches of the input matrix elements. Using an outer product-based SpGEMM kernel can reduce redundant fetches, but at the cost of increased overhead due to extra computation and memory accesses for producing/managing partial products. In this thesis, we introduce a novel SpGEMM kernel implementation based on the row-wise product approach. We leverage atomic instructions to merge intermediate partial products as they are generated. The use of atomic instructions eliminates the need to create partial product matrices, thus eliminating redundant DRAM fetches. To evaluate our row-wise product approach, we map an optimized SpGEMM kernel to a custom accelerator designed to accelerate graph-based applications. The targeted accelerator is an experimental system named PIUMA, being developed by Intel. PIUMA provides several attractive features, including fast context switching, user-configurable caches, globally addressable memory, non-coherent caches, and asynchronous pipelines. We tailor our SpGEMM kernel to exploit many of the features of the PIUMA fabric. This thesis compares our SpGEMM implementation against prior solutions, all mapped to the PIUMA framework. We briefly describe some of the PIUMA architecture features and then delve into the details of our optimized SpGEMM kernel. Our SpGEMM kernel can achieve 9.4x speedup as compared to competing approaches.

关键词： Computer science   Electrical engineering   Operations research   Military studies   Sustainability   Transportation  

摘要： The problem addressed in this study was the low counterfeit electronics parts prevention cost-efficiency for the U.S. legacy military aircraft sustainment, which was due in part to the lack of policy differentiation between legacy aircraft and other types of systems. The dissertation research utilized a qualitative before and after case study to explore potential cost-effectiveness improvements based on the use of a capable guardian as described in the Routine Activity Theory. Analysis of the current as-is case was based on existing U.S. laws and Department of Defense (DoD) regulations to identify themes and missing themes that could impact either performance or costs associated with the current counterfeit electronic parts prevention approach. The guardian-based after-case was qualitatively evaluated by DoD subject matter experts (SMEs). The SME opinions were gathered through the use of a survey. The study population was the set of senior Air Force engineers and computer scientists directly involved with the implementation of counterfeit parts prevention measures for fighter and attack legacy aircraft. No U.S. government systems or official time was used in the performance of the survey responses. The SME responses indicated that the use of a guardianship approach might improve counterfeit parts prevention cost-effectiveness within the DoD, if an appropriate implementation were utilized.

关键词： Applied mathematics   Mathematics   Computer science   Software packages   Equilibrium   Approximation   Numerical analysis   Methods   Algorithms   Geometry   Catastrophes   Efficiency   Linear algebra   Certification  

关键词： Computer science   Artificial intelligence   Information technology  

摘要： This thesis presents a novel architecture including two new programming languages, allowing software developers to create intelligent agents operating anywhere on the internet, independent of vendors' proprietary ecosystems or hardware. The open source architecture and accompanying languages, Huey and Jak, will allow developers to create vendor-independent, artificial intelligence solutions in the digital assistant space. Huey is a new programming language for building applications to allow the user to speak to or chat with a digital assistant using natural language. Huey uses a grammar-based approach to provide a natural language processing framework for turning user requests into actions. Jak provides an internal, intermediate code representation of actions and data. Jak will help add logic to secure conversations and protect people from inadvertently leaking emotion or private information to other parties. Developers can design and create their own voice and chatbot applications with custom grammars to carry out actions on mobile devices, in web pages, embedded devices, or on remote servers—independent of any particular operating system, cloud environment, or proprietary hardware.

关键词： Accuracy   Deep learning   Datasets   Voice recognition   Classification   Acoustics   Sound   Robotics   Computer science  

摘要： Although research on Acoustic Scene Classification (ASC) is very close to, or even overshadowed by different popular research areas known as Automatic Speech Recognition (ASR), Speaker Recognition (SR) or Image Processing (IP), this field potentially opens up several distinct and meaningful application areas based on environment context detection. The challenges of ASC mainly come from different noise resources, various sounds in real-world environments, occurring as single sounds, continuous sounds or overlapping sounds. In comparison to speech, sound scenes are more challenging mainly due to their being unstructured in form and closely similar to noise in certain contexts. Although a wide range of publications have focused on ASC recently, they show task-specific ways that either explore certain aspects of an ASC system or are evaluated on limited acoustic scene datasets. Therefore, the aim of this thesis is to contribute to the development of a robust framework to be applied for ASC, evaluated on various recently published datasets, and to achieve competitive performance compared to the state-of-the-art systems. To do this, a baseline model is firstly introduced. Next, extensive experiments on the baseline are conducted to identify key factors affecting final classification accuracy. From the comprehensive analysis, a robust deep learning framework, namely the Encoder-Decoder structure, is proposed to address three main factors that directly affect an ASC system. These factors comprise low-level input features, high-level feature extraction methodologies, and architectures for final classification. Within the proposed framework, three spectrogram transformations, namely Constant Q Transform (CQT), gammatone filter (Gamma), and log-mel, are used to convert recorded audio signals into spectrogram representations that resemble two-dimensional images. These three spectrograms used are referred to as low-level input features. To extract high-level features from spectr

关键词： Data processing   Geophysics   Historic artifacts   Software   Geology   Computer science  

摘要： The variable contexts of Fremont habitation sites in Utah Valley often make identification of those sites very challenging for archaeologists. Pit houses and other structures throughout the valley are frequently in plowed fields or other disturbed contexts that obscure their more exact location and nature. The application of geophysical technologies at archaeological sites throughout the world, including in North America, has proven to be an effective means of subsurface archaeological survey. However, geophysical techniques have been underutilized in Fremont archaeology. This paper reports on the employment of two geophysical methods, ground-penetrating radar (GPR) and fluxgate gradiometer surveys, at three known Fremont habitation sites in southern Utah Valley – the Wolf Village, Wolf Mound, and Snow Farm sites. The preliminary geophysical surveys and later ground-truthing of various geophysical anomalies reveals the effectiveness of these methods in identifying where architectural or other cultural features exist below the surface.

关键词： Computer science   Control theory   Teaching   Poisoning   Dynamic programming   Artificial intelligence   Experiments   Algorithms   Dynamical systems   Boundary value problems   Poisons   Distance learning  

摘要： This thesis provides an overview of recent results in adversarial online learning due to myself and my collaborators. The key question is the following: Consider an online learning setting, e.g. online supervised learning, bandits, or reinforcement learning, in which we define 3 entities: the underlying environment, the adversary, and the learning agent. During the learning process, the agent tries to achieve the learning goal by interacting with the environment, whereas the adversary desires to mislead the learner to achieve an attack goal by contaminating the interaction between the agent and the environment. This general setting of adversarial learning can be viewed as a two-player game between the attacker and the learner. The learner plays first by choosing a learning strategy. Observing the learning strategy and the environment, the attacker plays second by choosing an attack strategy. Notice that there is an information gap between the attacker and the learner. A (white-box) attacker observes both the strategy of the learner and the mechanics of the environment and therefore can find the corresponding optimal strategy directly. On the other hand, the learner observes neither the environment's mechanics (otherwise there is no need to learn) nor the attacker's strategy/utility, and therefore has no better choice than deploying a min-max strategy, hoping to maximize its utility against the worst-case adversary. Traditionally, this problem has been extensively studied in the offline learning setting. Taking supervised learning as an example, where the environment generates a pool of i.i.d. data from the underlying distribution D. The attacker has the power to contaminate the data by injection/deletion/modification, and the learner must perform learning on the contaminated data. Prior work has studied this problem from both the attacker's side and the learner's side. On the attacker's side, this is sometimes referred to as the data poisoning problem. It is shown t

关键词： Computer science   Artificial intelligence  

摘要： Deep Learning has been considered as one of the most significant milestones for Artificial Intelligence. Deep convolutional neural networks (CNNs) have been proven to be very successful in many pattern recognition tasks. Used as a pre-trained base network trained on powerful computers on large datasets, CNNs offer remarkable transfer learning capabilities. The CNN features learned in a local-to-global pyramidal architecture extracts gradually more sophisticated features in the higher layers based on, the lower ones' features. The hidden layers' connection weights provide broad-purpose (pluripotent) features that can be transferred to other networks for new target tasks, such as recognizing new object classes with possibly smaller datasets. Supervised deep CNNs are trained top-down and minimizing the classification error on large manually labeled datasets with thousands of classes, thus achieving their learning of pluripotent features. Recently proposed Contextually Guided Convolutional Neural Network (CG-CNN) architecture learns to extract such pluripotent features of a single convolutional layer in an unsupervised setting. Although CG-CNN has an advantage over deep CNNs, it can tune to these pluripotent features bottom-up without requiring massive, labeled datasets; its semi-supervised and multi-layered extensions require further research. This thesis proposed an extension named BeiimNet and demonstrated its effectiveness in applying the CG-CNN principles to semi-supervised complex pattern recognition tasks.