教学课程资源库 更多>>

关键词： Computer science   Computer engineering   Systems science  

摘要： State-of-the-art distributed deep learning systems, such as TensorFlow and PyTorch, are built on rigid assumptions that tightly couple model training and inference with the underlying hardware. First, they assume resource allocations must be fixed throughout the lifetime of a job, often leading to inefficient resource usage. Second, they require model hyperparameters to be retuned across different hardware configurations in order to achieve the same training result, posing a significant burden on the user. Due to these requirements, users are forced to juggle both systems challenges and application logic instead of being able to focus on just the latter. In this dissertation, we demonstrate that the above assumptions are not fundamental to distributed deep learning. We resolve these limitations by proposing two systems built on top of TensorFlow. The first is an autoscaling engine that, through trial-and-error, automatically determines the most resource-efficient hardware configuration for a given job. We propose pluggable heuristics tailored for deep learning workloads that incrementally guide the system towards such a configuration. Instead of repeatedly stopping the job and restarting it from checkpoints, which can lead to expensive hardware accelerators (e.g. GPUs, TPUs) going idle for minutes every time, our system adjusts the job’s all-reduce membership dynamically in between training steps without interrupting the job. The second system is VirtualFlow, which leverages a novel abstraction between the model and the underlying hardware called virtual node processing. From the perspective of the model, virtual nodes, instead of physical hardware accelerators, perform the computation. When multiple virtual nodes are mapped to a physical device, they are processed sequentially on that device. This representation offers users the flexibility to trade off computation time with resource requirement, allowing them to train their models using the same sets of hyperparam

关键词： Computer science   Artificial intelligence  

关键词： Computer science   Artificial intelligence   Java   Datasets   Graphs   Experiments   Benchmarks   Mathematical functions   Data base management systems   Connectivity   Arrays   Algorithms   Libraries   Linear algebra  

摘要： Graph Frameworks and Databases are critical components of modern software. The need to analyze massive graph datasets have spurred the development of Graph Systems. Graph Frameworks and Libraries with tuned Graph data structures are continuously being developed to handle new workloads and data patterns. This presents a need for a Graph system that has knowledge of its design space and is capable of combining fundamental design constructs to generate optimal graph data structures for a given hardware, data pattern and workload. We propose leveraging non-graph systems with these capabilities and with an overlap in its design space with Graph systems, to bring this intelligence to Graph Systems. As a first step in this process, we have implemented a Key-Value Graph Generator, that demonstrates the use of key-value approach in designing Adjacency List and Compressed Sparse Row (CSR). Our hypothesis is that if we can successfully model Graph data structures using key-value approach then we can leverage learned key-value system and create an interactive and automatic Graph system.

关键词： Aerospace engineering   Electrical engineering   Computer science  

摘要： The purpose of this project was to improve an existing system currently in use by NASA Langley Research Center (LaRC). The 6-inch Magnetic Suspension and Balance System (MSBS) built at MIT is operational with control in three degrees of freedom, with two additional degrees of freedom exhibiting passive stability. The means for measuring model displacement within the magnetic environment is an Electromagnetic Position Sensor (EPS), consisting of excitation coils at 20 kHz and multiple sets of pickup coils. The pickup coil voltages are proportional to model displacement in each degree of freedom. However, the EPS electronic signal processing system is analog and outdated; setup and adjustment are time consuming. The task was to construct a one degree of freedom model of the EPS including its electronics system in order to explore digital signal processing. The model core is allowed only axial displacement for simplicity; this model coil set is essentially a Linear Variable Differential Transformer (LVDT). The source signal was chosen to be 2.36 kHz for convenience, with scaling up to the full size system at 20 kHz possible. An amplification and filtering circuit board was constructed to modify the signal for the proper functionality of the model EPS. By comparing the reference or excitation coil signal and the measured or pickup coil signal by means of a digital phase measurement method using cross-correlation analysis, the digital algorithm resolves the phase shift between the two signals and their amplitude ratio. The key proof of concept is the digital signal processing algorithm; since the defining characteristics of any signal are the amplitude and phase this algorithm can be adapted to suit various control and setup needs of the MSBS and EPS. By proving that a digital interface with the EPS is possible, the analog interface can be replaced with a digital system.

关键词： Computer engineering   Information technology   Computer science   Artificial intelligence  

摘要： The Internet has become a necessity in today's digital world, and making it secure is a pressing concern. Hackers are investing ever-increasing efforts to compromise Internet nodes with novel techniques. According to Forbes, every minute, $ 2,900,000 is lost to cybercrime. A common cyber-attack is Denial of Service (DoS) or Distributed Denial of Service (DDoS) attacks, which may bring a network to a standstill, and unless mitigated, network services could be halted for an extended period. The attack can occur at any layer of the OSI model. This thesis focuses on SYN Flood DoS/DDoS attacks, also known as TCP Flood attacks, and studies the use of artificial neural networks to detect the attacks. Specific neural network models used in this thesis are the Gated Recurrent Units (GRU), the Long Short-Term Memory (LSTM), and a semi-supervised model on label propagation. All neural network models detect attacks by analyzing the individual hexadecimal values in the packet header. A novelty of the approach followed in this thesis is that the neural networks do not consider the lexical values of the network packet (MAC addresses, IP addresses, and port numbers) as input features in their traffic analysis. Instead, the neural network models are designed and trained to detect malicious traffic based on the time pattern of TCP flags. The neural networks base their analysis of traffic on time-sequenced patterns. An important hyperparameter discussed in this paper is the size of the lookup window, that is, the number of past packets the model can access to predict the next packet. Evaluation results based on datasets presented in this thesis show that the accuracies of the GRU, CNN/LSTM, and label propagation models are 81%, 93%, and 96%, respectively.

关键词： Biomedical engineering   Computer science   Artificial intelligence   Biophysics  

摘要： Human activity recognition (HAR) is the process of predicting different human activity states (walking, running, walking, etc.) based on sensor data. HAR is mainly used in healthcare for geriatric activity monitoring and human machine interaction. Recurrence plot (RP) is visualization of a non-linear time series in a matrix form where the matrix elements correspond to the recurrence of a dynamical state. Recurrent Quantification Analysis (RQA) is a popular method to extract features from recurrence plot that can be used to represent the underlying dynamics of the time series. RQA quantifies the number and duration of recurrences of a dynamical system presented by its phase space trajectory. Human activity monitoring generates continuous sensor data and RQA computation is computationally expensive for long time series. To address this issue, in this work we investigate the effectiveness of the image classification method to classify a given recurrence plot. We use real-world human activity data to demonstrate the proposed approach. The activities from human subjects are measured by accelerometers and gyroscopes positioned at 7 different body locations (head, forearm, upper arm, chest, thigh, shin, and waist). For the classification task, we select running, walking, climbing up, and climbing down. The sensor signals from the accelerometer and gyroscope are captured for the aforementioned human activities. These sensor signals are converted into recurrence plots. The threshold value is set as 0.5 and 0.8 for generating the recurrence plots. The size of the recurrence plot is set to 32x32 pixels. Then a Convolutional Neural Network (CNN) is used to classify the human activities using the recurrence plots. A LeNet architecture is selected for the deep learning process. Recurrence plots of different human activities are fed into the LeNet5 algorithm in the form of epochs, with a batch size of 256 plots. We found that among all the sensor positions, the sensors fixed on sh

关键词： Engineering   Electrical engineering   Computer science  

摘要： Next generation cellular networks are expected to support a massive data traffic volume and satisfy a vast number of users that have latency-critical quality-of-service expectations. Towards serving this demand, it is envisaged that the interference management problem will be the main bottleneck due to the likeliness of a heavily interfering wireless environment caused by much denser deployment of base stations and mobiles. Due to inherently scarce shared frequency-band resources over time-varying traffic and multi-channel conditions, a scalable and practical fast-timescale resource management is an absolute necessity towards next generation cellular networks. The conventional optimization based resource management schemes are either practically infeasible, computationally challenging, or intractable due to relying on model-driven techniques. Therefore, in the past few years, there has been extensive research on model-free reinforcement learning based resource management. Reinforcement learning is purely data-driven, and its multi-agent adaptation is promising for scalability on larger networks where agents collaboratively work together towards a shared objective. We initially show the potential of deep reinforcement learning for transmit power control in wireless networks. Existing power control techniques typically find near-optimal power allocations by solving a challenging optimization problem. Most of these algorithms are not scalable to large networks in real-world scenarios because of their computational complexity and instantaneous cross-cell channel state information (CSI) requirement. The proposed method is a distributively executed dynamic power allocation scheme that maximizes a weighted sum-rate objective, which can be particularized to achieve maximum sum-rate or proportionally fair scheduling. Each transmitter collects delayed channel measurements of a time-varying channel from its neighbors and adapts its own transmit power accordingly. Both random v

关键词： Computer science   Artificial intelligence  

摘要： Unit tests are a key component during the software development process, helping ensure that a developer's code is functioning as expected. Developers interact with unit tests when trying to understand, maintain, and when updating code. Good test names are essential for making these various processes easier, which is important considering the substantial costs and effort of software maintenance. Despite this, it has been found that the quality of test code is often lacking, specifically when it comes to test names. When a test fails, its name is often the first thing developers will see when trying to fix the failure, therefore it is important that names are of high quality in order to help with the debugging process. The objective of this work was to find anti-patterns having to do with test method names that may have a negative impact on developer comprehension. In order to do this, a grounded theory study was conducted on 12 open-source Java and C# GitHub projects. From this dataset, many patterns were discovered to be common throughout the test code. Some of these patterns fit the necessary criteria of anti-patterns that would probably hinder developer comprehension. With the avoidance of these anti-patterns it is believed that developers will be able to write better test names that can help speed the time to debug errors as test names will be more comprehensive.

关键词： Algorithmic game theory   Congestion games   Equilibrium computation   Game theory   Multi-sided markets  

摘要： This dissertation considers a number of problems in pure and applied game theory. The first chapter considers the problem of how the introduction of fines and monitoring affects welfare in a routing game. I characterize equilibria of the game and discuss network topologies in which the introduction of fines can harm those agents which are not subject to them. The second, and primary, chapter considers the computational aspects of tenable strategy sets. I characterize these set-valued solution concepts using the more familiar framework of perturbed strategies, introduce strong alternatives to the problems of verifying whether a strategy block satisfies the conditions of tenability, and provide some hardness results regarding the verification of fine tenability. Additionally, I show an inclusion relation between the concept of coarse tenability and the notion of stability introduced by Kohlberg and Mertens (1986). Finally, I show how the methods developed for tenability provide an alternative characterization for proper equilibria in bimatrix games. This characterization gives a bound on the perturbations required in the definition of proper equilibria, though such bounds cannot be computed efficiently in general. The third, and final, chapter develops a model of contracting for content creation in an oligopolistic environment of attention intermediaries. I characterize symmetric equilibria in single-homing (exclusive) and multi-homing regimes. The focus is on the trade-off between reductions in incentives offered by intermediaries and the benefits of access to additional content for consumers. I show that when the extent of multi-homing is exogenous in the absence of exclusivity clauses, consumer surplus is always higher with multi-homing than under exclusivity, despite weaker incentives offered by platforms to content creators.

关键词： Electrical engineering   Computer science   Artificial intelligence  

摘要： The progression of deep neural networks (DNN) has allowed for new solutions to control problems in power electronics. Many works have supplemented controllers with DNNs in offline-learning, often verified in simulation. However, actual implementation of neural network intelligence for real-time dynamic systems has not been extensively investigated. While GPUs are used for processing in many data applications, their potential for electrical control systems could be explored. In this paper, a real-time dynamic system identification using a DNN implemented in a GPU on an embedded-Linux single-board computer is presented. A second-order analog circuit is used for the target system and relative analog/digital signal interfaces and Linux/CUDA firmware have been developed. The proposed system identified the system characteristics well in simulation. These results were then realized in an experimental hardware setup, demonstrating good computation speed.