教学课程资源库 更多>>

关键词： Information science   Artificial intelligence   Computer science  

摘要： Document classification is abstract task in the domain of natural language processing and information retrieval. There are traditional methods associated with this task, our method shows the performance enhancement in terms of the performance, convergence and enrichment of information. We propose a hybrid neural language modeling architecture that constructs hierarchical feature representations. We examine our architecture through document classification. In our first model we begin with a character level convolutional neural layer (CNN) to get word level representation, next layers recurrent neural network (RNN) with attention based feature merging in order to get sentence level representation and again we have RNN with attention layer to get document level representation and finally, we have interconnected dense structure stacked to classify documents with soft-max activation. We extend this model to the word level and summarize the overall results and comparisons with baseline models. We show evidence of the hypotheses on multiple datasets, utilizing IMDB YELP review datasets. We show extended results with all datasets in terms of performance with F1 score, accuracy, precision and recall. Also, we show the comparison of convergence time and rate of convergence of our approach. Moreover, we show visual evidence that our approach lead to better feature construction and able to construct features for 99% of the effective word vocabulary from the characters in the documents.

关键词： Computer science  

摘要： Experts predict that machine learning will become a commodity in the future. Recent research, however, have shown vulnerabilities of ML models served on the cloud, among them are adversarial attacks that perturb the trained parameters to influence the model output. Consider a deep neural network served on the cloud whose parameters have been tweaked. Can we solve for a sample that is sensitive to this tweaking thus expose the violation? In this study we use SMT solvers to retrieve sensitive samples. We devise three approaches to retrieve them: we find one about maximum sensitivity, or we specify the conditions of the attack to the solver, or we transform a training input become a sensitive sample. Among our analyses, samples within the neighborhood of our optimal sensitive sample are still good candidates for detection. Another, sensitive samples can detect any actual weight perturbation for as long it is active on that neuron. Furthermore, we motivate a strategy for very large DNNs where we patch only a subset of the inputs as free variables while the rest are constants. Our efforts toward securing these models are sure contribution to society today and tomorrow.

关键词： Information science   Computer science  

摘要： Data mining and machine learning equipped with large-scale data is an interesting and popular research topic and in both academia and industry. Leveraging large data has been shown helpful for a wide variety of problems and application, and it is still in active development. Large data introduces challenges to almost all aspects of data mining and machine learning, including performance, efficiency, scalability, deployment, etc. For this dissertation, we focus on the relational data and develop a community-based approach to develop a framework capable of handling large-scale relational data by the “divide-and-conquer” strategy, while we also involve techniques like pre-training and sampling to improve performance generality and scalability. Relational data is ubiquitous; it is not just the graph data with explicit links like social network or knowledge base; any data with feature vectors can have underlying relations by similarities or correlations. The topology of a set of real-life relational data tend to have underlying community-like structures of which we can take advantage. This thesis develops methods that try to solve sub-problems on partitions of a large dataset and then merge the results if necessary. We target two concrete problems, network alignment and correlated anomaly detection. We start with the probabilistic approach, based on a classic statistical model called the Stochastic Block Model (SBM), and an probabilistic embedding technique called the Latent Dirichlet Mixture. The design is specific for each problem. For network alignment, an approximate likelihood is found to work for real-life networks with a constant densification power, and reduce the overall alignment time complexity from quadratic to sub-quadratic. The experiments show our approach is superior to some recent models intended to handle large graphs. For correlated anomaly detection, we apply a variant of SBM by treating the correlation matrix as a weighted graph. We proved our graph-

关键词： Spatial resolution   Face   Image reconstruction   Machine learning   Image restoration   Computer science   Facial video hallucination   recurrent frame fusion   sequential feature encoding   deep learning  

摘要： As a domain-specific super-resolution problem, facial image hallucination has enjoyed a series of breakthroughs thanks to the advances of deep convolutional neural networks. However, the direct migration of existing methods to video is still difficult to achieve good performance due to its lack of alignment and consistency modelling in temporal domain. Taking advantage of high inter-frame dependency in videos, we propose a self-enhanced convolutional network for facial video hallucination. It is implemented by making full usage of preceding super-resolved frames and a temporal window of adjacent low-resolution frames. Specifically, the algorithm first obtains the initial high-resolution inference of each frame by taking into consideration a sequence of consecutive low-resolution inputs through temporal consistency modelling. It further recurrently exploits the reconstructed results and intermediate features of a sequence of preceding frames to improve the initial super-resolution of the current frame by modelling the coherence of structural facial features across frames. Quantitative and qualitative evaluations demonstrate the superiority of the proposed algorithm against state-of-the-art methods. Moreover, our algorithm also achieves excellent performance in the task of general video super-resolution in a single-shot setting.

关键词： Computer science  

摘要： As the Internet becomes increasingly crucial to distributing information, Internet censorship has become more pervasive and advanced. A common way to circumvent Internet censorship is Tor, a network that provides anonymity by routing traffic through various servers around the world before it reaches its destination. However, adversaries are capable of identifying and censoring access to Tor due to identifying features in its traffic. Meek, a traffic obfuscation method, protects Tor users from censorship by hiding Tor traffic inside an HTTPS connection to a permitted host. This approach provides a defense against censors using basic deep packet inspection (DPI), but machine learning attacks using side-channel information against Meek pose a significant threat to its ability to obfuscate traffic. In this thesis, we develop a method to 1. efficiently gather reproducible packet captures from both normal HTTPS and Meek traffic, 2. aggregate statistical signatures from these packet captures, and 3. train a generative adversarial network (GAN) to minimally modify statistical signatures in a way that hinders classification. Our GAN successfully decreases the efficacy of trained classifiers, increasing their mean false positive rate (FPR) from 0.183 to 0.834 and decreasing their mean area under the precision-recall curve (PR-AUC) from 0.990 to 0.414.

关键词： Computer engineering   Computer science  

摘要： Rapid and accurate identification of hospitalized patients at high risk for readmission, disease, extended length of stay (LOS), mortality, etc., has the potential to improve quality of care and reduce avoidable harm and costs. However, most data driven studies for risk prediction in a healthcare setting have produced non-interpretable black boxes, which precludes them from being used effectively within the decision support systems in the hospitals. The focus of this dissertation is on developing techniques to improve the interpretability and explainability of machine learning models in the context of healthcare by incorporating domain knowledge, specifically for the task of predicting the risk of hospital readmission. Preventable hospital readmissions have been identified as one of the primary targets for reducing costs and improving healthcare delivery, and advancements in data-driven approaches for this critical task can potentially have a significant impact on the healthcare system. Three approaches have been proposed in this dissertation to predict readmission risk. In the first two approaches, we focus on incorporating domain knowledge, in the form of hierarchical taxonomies available for disease codes, to improve the interpretability of a linear readmission prediction model. Two models are proposed with accuracies that are comparable to state of art methods. However, both models produce a highly interpretable output, which allows medical experts to draw clinically relevant insights and identify key factors associated with hospital readmissions. In both models, we exploit the domain induced hierarchical structure available for the disease codes which are the features for the classification algorithm. In the first approach, a structured sparsity regularization based model is applied. In the second approach, to improve the interpretability, a novel tree-structured sparsity-inducing regularization norm is proposed. Furthermore, a quantitative evaluation metric to

关键词： capital punishment   crime   criminal code   exclusiveness   hooliganism   life imprisonment   public danger   punishment   responsibility   исключительность   наказание   общественная опасность   ответственность   пожизненное лишение свободы   преступление   смертная казнь   уголовный кодекс   хулиганство   Language of Keywords: English   Russian  

摘要： В статье впервые предпринято системное исследование категории исключительно- сти в уголовном законодательстве и выявлена ее социально-правовая природа. Отмеча- ется, что в уголовном законодательстве речь может идти об исключительных условиях совершения преступлений, его признаках, усиливающих или смягчающих ответствен- ность, либо же об исключительных уголовно-правовых мерах. Проведен анализ осно- вания и условий применения смертной казни, а также проявлений исключительного цинизма при совершении хулиганства и исключительных обстоятельств как основания назначения более мягкого наказания, чем предусмотрено за данное преступление. Сде- лан вывод о том, что исключительность в уголовном праве может обусловливать как усиление, так и смягчение наказания. Поскольку речь в этих случаях идет о чрезвычай- ном усилении или смягчении наказания, необходимо исчерпывающим образом регла- ментировать в законе основания их применения. Кроме того, обоснованы предложения по совершенствованию некоторых норм Уголовного кодекса Российской Федерации. [ABSTRACT FROM AUTHOR]

关键词： Artificial intelligence   Computer science  

摘要： Socioeconomic status and gender are are heavily influenced by systemic bias. We focus on addressing systemic bias in the area of socioeconomic status by introducing an intelligent mathematical tutoring system, HWHelper. HWHelper is designed to be accessible by students, so that students who may not otherwise have help at home have access to guided instruction. The understand the systemic bias associated with gender, we introduce a framework for the identification and evaluation of gender bias in political news articles. We find that HWHelper performs well on the limited dataset. We find that our political gender bias detection system finds clear differences in the language used to describe male versus female political candidates.

关键词： Measurement   Computer science   Electronic mail   Standards   Bridge circuits   Organizations   Licenses   Closed trail distance   clustering coefficient   closure coefficient   cyclic structure   higher-order structure  

摘要： Real-world networks contain many cliques since they are usually built from them. The analysis that goes behind the cliques is fundamental because it discovers the real structure of the network. This article proposed new high-order closed trail clustering and closure coefficients for evaluation of the network structure. These coefficients are able to describe the inner structure of the network concerning its randomized or organized behavior. Moreover, the coefficients can cluster networks with similar structures together. The experiments show that the coefficients are useful in both the local and global context.

关键词： Information Science   Information Technology   Information Systems   Computer Science   Legal Studies   Systems Design   information governance   records management   compliance   audits   auditing   policy risk   security  

摘要： Audits are rarely used by organizations to effectively improve their information governance (IG) programs. The objective of this qualitative, semi-structured interview research was to determine if an organization that performed an audit of their IG program could improve compliance and policy objectives. The sample included four participants from Federal, state, and municipal government agencies that met level a three of ARMA's GARP Maturity Model (2012) regarding an organization’s regulatory, compliance, and policy requirements. Participants conducted at least one audit of their IG program over an eight-year period. The findings concluded that the utilization of audits improved IG program policy compliance and mitigated information risk in an organization. Implications and recommendations are discussed.