教学课程资源库 更多>>

关键词： Computer science   Artificial intelligence   COVID-19  

摘要： The proliferation of digital technologies has led to an explosion in the number of large datasets available in the last few years, placing traditional machine learning approaches to data processing and modeling at a competitive disadvantage. Nevertheless, analyzing complex, high-dimensional, and noisy datasets can be a tremendous challenge. Deep learning, as part of a broader family of machine learning methods, has shown superior performance in dealing with such challenges in the past decade. However, several challenges in the deep learning lifecycle hinder the performance and democratization of deep learning methods. This dissertation spotlights a key challenge: efficiency. Specifically, we focused on three topics: efficient representation learning, efficient temporal model learning, and efficient model compression. The three topics correspond to the sequential stages of data representation, modeling, and deployment in the deep learning lifecycle. The first topic is efficient representation learning. Our research focuses on the field of knowledge graph representation learning. Though researchers have investigated several knowledge graph embedding methods, efficiently comparing them with existing solutions and exploring new ones remains challenging. We have, thus, proposed a unified group-theoretic framework for general knowledge graph embedding problems and explored two novel efficient embedding methods that though compact in size, demonstrate impressive results on benchmark datasets. The second topic is efficient temporal model learning. As a significant part of artificial intelligence, temporal learning utilizes temporal data to predict future events or infer latent traits. We found that numerous deep learning methods are focused on computer vision and natural language processing though efficient prediction models for temporal learning are in demand. This thesis proposes three efficient prediction models in temporal learning that can deliver superior performance

关键词： Computer science   Biology   Medical imaging   Ecology   Datasets   Neural networks   COVID-19   Medical research  

摘要： Deep learning in computer vision and image processing has attracted attentions from various fields including ecology and medical image. Ecologists are interested in finding an effective model structure to classify different species. Tradition deep learning model use a convolutional neural network, such as LeNet, AlexNet, VGG models, residual neural network, and inception models, are first used on classifying bee wing and butterfly datasets. However, insufficient data sample and unbalanced samples in each class have caused a poor accuracy. To make improvement the test accuracy, data augmentation and transfer learning are applied. Recently developed deep learning framework based on mathematical morphology also shows its effective in shape representation, contour detection and image smoothing. The experimental results in the morphological neural network shows this type of deep learning model is also effective in ecology datasets and medical dataset. Compared with CNN, the MNN could achieve a similar or better result in the following datasets. The chest X-ray images are notoriously difficult to analyze for the radiologists due to their noisy nature. The existing models based on convolutional neural networks contain a giant number of parameters and thus require multi-advanced GPUs to deploy. In this research, the morphological neural networks are developed to classify chest X-ray images, including the Pneumonia Dataset and the COVID-19 Dataset. A novel structure, which can self-learn a morphological dilation or erosion, is proposed for determining the most suitable depth of the adaptive layer. Experimental results on the chest X-ray dataset and the COVID-19 dataset show that the proposed model achieves the highest classification rate as comparing against the existing models. More significant improvement is that the proposed model reduces around 97% computational parameters of the existing models. Automatic identification of pneumonia on medical images has attracted inten

关键词： Computer science   Natural Language Generation   Reinforcement Learning   Sequential Decision-making   Uncertainty Estimation  

摘要： Deep reinforcement learning (RL) has received great success in playing video games and strategic board games, where a simulator is well-defined, and massive samples are available. However, in many real-world applications, the samples are not easy to collect, and the collection process may be expensive and risky. We consider designing sample efficient RL algorithms for online exploration and learning from offline interactions. In this thesis, I will introduce algorithms that quantify uncertainty via exploiting intrinsic structures within observations to improve sample complexity. These proposed algorithms are theoretically sound and show broad applicability in recommendation, computer vision, operations management, and natural language processing. This thesis consists of two parts: (i) efficient exploration and (ii) data-driven reinforcement learning. Exploration-exploitation has been widely recognized as a fundamental trade-off. An agent can take exploration actions to learn a better policy or take exploitation actions with the highest reward. A good exploration strategy can improve sample complexity as a policy can converge faster to near optimality via collecting informative data. Better estimation and usage of uncertainty lead to more efficient exploration, as the agent can efficiently explore to better understand environments, \textit{i.e.}, minimizing uncertainty. In the efficient exploration part, we place the reinforcement learning into the probability measure space and formulate it as Wasserstein gradient flows. The proposed method can quantify the uncertainty of value, policy, and constraint functions to provide efficient exploration. Running a policy in real environments can be expensive and risky. Besides, there are massive logged datasets available. Data-driven RL can effectively exploit these fixed datasets to perform policy improvement or evaluation. In the data-driven RL part, we consider auto-regressive sequence generation as a real-world sequential

关键词： Stock markets   Companies   Pandemics   Voting   Time series analysis   Correlation   Computer science   Event database   social media data   sentiment analysis   data analytics   event studies  

摘要： An event is a piece of news that triggers a change in stock prices. Here, an event study is undertaken to capture the effect of abnormal returns due to an event. The event can affect the stock market in the long term or short term. Event research is relevant to both the efficient market hypothesis and behavioral finance. In this study, we collected data from websites that manage financial and economic data, performed a sentiment analysis, and correlated news article data with changes in a particular company's stock prices in the stock market. Data were collected from two well-known financial news websites. We observed a correlation between stock prices and news items. An event period of one day was considered for the study. The regression equation determined the relationship between stock returns and polarity and subjectivity. Bayesian model averaging was performed to identify the effects of polarity and subjectivity on stock returns. Time-series data were decomposed into components and detrended via regression. Prominent keywords and their polarity values for a particular day were plotted. An event enhanced some stock returns while adversely affecting other stocks. We found a variation range of -400 to 200 for different company stocks for the selected period.

关键词： Mathematics   Computer science   Physics   Eigen values   Algebra   Quantum theory   Algorithms   Information theory   Expected values  

摘要： This work extends the theories of algorithmic randomness and Kolmogorov complexity of bitstrings to the quantum realm. Nies and Scholz defined quantum Martin-Löf randomness (q-MLR): the first notion of algorithmic randomness to be defined for infinite sequences of qubits, which are called states. We define a notion of quantum Solovay randomness and show it to be equivalent to q-MLR using purely linear algebraic methods. Quantum Schnorr randomness is then introduced. A quantum analogue of the law of large numbers is shown to hold for quantum Schnorr random states. We next turn to a quantum analogue of Kolmogorov complexity. We introduce quantum-K (QK), a measure of the descriptive complexity of density matrices using classical prefix-free Turing machines and show that the initial segments of weak Solovay random and quantum Schnorr random states are incompressible in the sense of QK. Many properties enjoyed by prefix-free Kolmogorov complexity (K) have analogous versions for QK; notably a counting condition. Several connections between Solovay randomness and K, including the Chaitin type characterization of Solovay randomness, carry over to those between weak Solovay randomness and QK. Schnorr randomness has a Levin–Schnorr characterization using KC; a version of K defined using an arbitrary computable measure machine, C. We similarly define QKC, a version of QK. Quantum Schnorr randomness is shown to have a Levin–Schnorr and a Chaitin type characterization using QKC. We then show how classical randomness can be generated from a computable, non-quantum random state. We formalize how `measurement' of a state induces a probability measure on the space of infinite bitstrings. A state is 'measurement random' (mR) if the measure induced by it, under any computable basis, assigns probability one to the set of Martin-Löf randoms. I.e., measuring a mR state produces a Martin-Löf random bitstring with probability one. While quantum-Martin-Löf random states are mR, we show that

关键词： Computer engineering   Computer science   Artificial intelligence  

摘要： Neural networks are increasingly becoming attractive solutions for automated systems within automotive, aerospace, and military industries. Since many applications in such fields are both real-time and safety-critical, strict performance and reliability constraints must be considered. To achieve high performance, specialized architectures are required. Given that over 90% of the workload in modern neural network topologies is dominated by matrix multiplication, accelerating said algorithm becomes of paramount importance. Modern neural network accelerators, such as Xilinx’s Deep Processing Unit (DPU), adopt efficient systolic-like architectures. Thanks to their high degree of parallelism and design flexibility, Field-Programmable Gate Arrays (FPGAs) are among the most promising devices for speeding up matrix multiplication and neural network computation. However, SRAMbased FPGAs are also known to suffer from radiation-induced upsets in their configuration memories. To achieve high reliability, hardening strategies must be put in place. However, traditional modular redundancy of inherently expensive modules is not always feasible due to limited resource availability on target devices. Therefore, more efficient and cleverly designed hardening methods become a necessity. For instance, Algorithm-Based Fault-Tolerance (ABFT) exploits algorithm characteristics to deliver error detection/correction capabilities at significantly lower costs. First, experimental results with Xilinx’s DPU indicate that failure rates can be over twice as high as the limits specified for terrestrial applications. In other words, the undeniable need for hardening in the state-of-the-art neural network accelerator for FPGAs is demonstrated. Later, an extensive multi-level fault propagation analysis is presented, and an ultralow-cost algorithm-based error detection strategy for matrix multiplication is proposed. By considering the specifics of FPGAs’ fault model, this novel hardening method decreas

关键词： Computer science   Artificial intelligence  

摘要： Machine learning (ML) has come to be widely used in a broad array of settings, including important security applications such as network intrusion, fraud, and malware detection, as well as other high-stakes settings, such as autonomous driving. A general approach is to extract a set of features, or numerical attributes, of entities in question, collect a training data set of labeled examples (for example, indicating which instances are malicious and which are benign), learn a model which labels previously unseen instances presented in terms of their extracted features, and then investigate alerts raised by instances predicted as malicious. Despite the striking success of ML in security applications, security issues emerge from the full pipeline of ML-based detection systems. First, ML models are often susceptible to adversarial examples, in which an adversary makes changes to the input (such as malware) to avoid being detected. Second, using detection systems in practice is dealing with an overwhelming number of alerts that are triggered by normal behavior (the so-called false positives), obscuring alerts resulting from actual malicious activities. Third, adversaries can target a broad array of ML-based detection systems to maximize impact, which is often ignored by individual ML system designers. In this thesis, I focus on studying the security problems of deploying robust machine learning systems in adversarial settings. To conduct systematic research on this topic, my study is based on four components. First, I study the problem of systematizing adversarial evaluation. Concretely, I propose a fine-grained robustness evaluation framework for face recognition systems. Second, I investigate robust machine learning against decision-time attacks. Specifically, I propose a framework for validating models of ML evasion attacks, and evaluate the efficacy of conventional robust machine learning models against realizable attacks in PDF malware detection. My work shows that

关键词： Software   Physical chemistry   Partial differential equations   Chemical reactions   Numerical analysis   Algebra   Dynamical systems   Boundary conditions   Boundary value problems   Applied mathematics   Chemistry   Computer science   Mathematics  

关键词： Computer science   Artificial intelligence  

摘要： Generative models can serve as a powerful primitive for creative interaction with data. Generative models give us the ability to synthesize or re-synthesize multimedia; conditional generative modeling empowers us to control the outputs of these models. By steering a generative model with conditioning information, we can sketch the essential aspects of our creative vision, and the generative model will fill in the details. This dissertation explores the possibilities of generative modeling as a creative tool, with a focus on applications to music and audio. The dissertation proceeds in three parts: 1. We de velop algorithms and evaluation metrics for aligning musical scores to audio. Alignments provide us with a dense set of labels on musical audio, that we can use to supervise conditional generation tasks such as transcription: synthesis of a musical score conditioned on an audio performance. This work on alignments leads to the construction of MusicNet: a collection of 330 freely-licensed classical music recordings, together with over 1 million annotated labels indicating the precise time of each note in every recording, the instrument that plays each note. We use this dataset to train state-of-the-art music transcription models for the MIREX Multiple Fundamental Frequency Estimation task. 2. We construct autoregressive generative models of musical scores, which exploit invariances in the structure of music. Whereas most recent work on music modeling has represented music as an ordered sequence of notes, we explore an alternative representation of music as a multi-dimensional tensor. We consider a variety of factorizations of the joint distribution over these tensors. We then turn to our attention to discriminative modeling of scores, using this tensor representation. We construct a classifier that can reliably identify the composer of a classical musical score. Our methods, which operate on the generic tensor score representation, outperform previously reported re

关键词： Switches   Control systems   IP networks   Servers   Topology   Protocols   Computer science   SDN   OpenFlow   in-band control   bootstrapping  

摘要： An SDN switch newly added to a network needs to establish a control channel with an SDN controller to manage control-plane traffic. The setup of such a channel is termed bootstrapping. In the case of in-band control, bootstrapping involves setting up a control path that may traverse one or more switches between the new switch and the controller, which is achieved through the configuration of layer-two to layer-four parameters on relevant switches. Previous approaches either result in lengthy bootstrapping time due to their level-by-level mechanisms, demand complicated modules, or lack some essential features. This paper proposes an approach to fast automatic bootstrapping that overlaps the bootstrapping processes among a set of switches. Our emulations using virtual switches completed the process of bootstrapping 50 switches arranged in a chain in 2.0 seconds, which is a 98% time reduction compared with a prior study.