教学课程资源库 更多>>

关键词： Computer science   Linguistics   Code switching (Linguistics)   Corpora (Linguistics)   Computational linguistics   Parts of speech  

摘要： Code-switching is the phenomenon by which bilingual speakers switch between multiplelanguages during written or spoken communication. The importance of developing lan-guage technologies that are able to process code-switched language is immense, given thelarge populations that routinely code-switch. Current NLP and Speech models break downwhen used on code-switched data, interrupting the language processing pipeline in back-endsystems and forcing users to communicate in ways which for them are unnatural. There are four main challenges that arise in building code-switched models: lack of code-switched data on which to train generative language models; lack of multilingual languageannotations on code-switched examples which are needed to train supervised models; littleunderstanding of how to leverage monolingual and parallel resources to build better code-switched models; and finally, how to use these models to learn why and when code-switchinghappens across language pairs. In this thesis, I look into different aspects of these fourchallenges. The first part of this thesis focuses on how to obtain reliable corpora of code-switchedlanguage. We collected a large corpus of code-switched language from social media using acombination of sets of anchor words that exist in one language and sentence-level languagetaggers. The newly obtained corpus is superior to other corpora collected via differentstrategies when it comes to the amount and type of bilingualism in it. It also helps trainbetter language tagging models. We also have proposed a new annotation scheme to obtainpart-of-speech tags for code-switched English-Spanish language. The annotation scheme iscomposed of three different subtasks including automatic labeling, word-specific questionslabeling and question-tree word labeling. The part-of-speech labels obtained for the MiamiBangor corpus of English-Spanish conversational speech show very high agreement and accuracy. The second section of this thesis focuses on the

关键词： Computer science  

摘要： Unknown regulatory motif finding on DNA sequences is a crucial task for understanding gene expression and the task requires accuracy and efficiency. We propose DMF, a combinatorial approach that uses hash-based heuristics to skip unnecessary computations while retaining the maximum accuracy. Parallelized versions of our DMF approach, called PDMF, have been developed to use CPU, GPU, and heterogeneous computing architectures in order to achieve the maximum performance. PDMF also incorporates SIMD instructions to further accelerate the task of unknown motif search. Our experimental results show that the multicore version (PDMFm) achieved 8.87x speedup over DMF. The GPU version (PDMFg) achieved a 41.48x and 9.95x average speedup over the serial version and PDMFm, respectively. Our SIMD enhanced heterogeneous approach (PDMFh) achieved a 3.42x speedup over our fastest GPU model (PDMFg1). The proposed approach was tested for performance against popular approximate and suffix tree-based approaches with various sized real-world datasets and the experimental results showed that the proposed approach achieved the maximum accuracy within a practical time bound for motif lengths 6~14.

关键词： Engineering   Artificial intelligence   Computer science  

摘要： In this thesis, we have proposed efficient implementations of linear algebra kernels such as matrix-vector and matrix-matrix multiplications by formulating arithmetic calculations in terms of diagonals and thereby giving an orientation-neutral (column-/row-major layout) computational scheme. Matrix elements are accessed with stride-1 and no indirect referencing is involved. Access to the transposed matrix requires no additional effort. The proposed storage scheme handles dense matrices and matrices with special structures such as banded, symmetric in a uniform manner. Test results from numerical experiments with OpenMP implementation are promising. We also show that, using our diagonal framework, Java native arrays can yield superior computational performance. We present two alternative implementations for matrix-matrix multiplication operation in Java. The results from numerical testing demonstrate the advantage of our proposed methods.

关键词： Computer science  

摘要： Land cover mapping is the task of generating maps of land use globally across time. The recent decades have seen an increasing availability of public satellite data sets with observations of the Earth at regular intervals of space and time. This coupled with the advances in machine learning and high performance computing provide an opportunity to automate the land cover mapping problem at scale. However, the availability of labeled data to train predictive models in this application is very limited, especially in the developing regions of the world, where accurate land cover maps are necessary for effective management of natural resources to sustain the rapid population growth in these regions. The need for labeled samples is further increased by: (1) Heterogeneity of land cover classes across space and time; (2) Increasing complexity of state-of-the-art predictive models and (3) Lack of sufficient samples at the required spatial and temporal resolutions. Since paucity of labeled data is a major problem in this domain, traditional machine learning algorithms that only rely on exact labeled data (strong supervision) have limited performance. This thesis investigates the use of weak supervision to mitigate the problem of not having sufficient samples with exact labels. In a weakly-supervised learning scenario, you have very few training samples that have exact labels corresponding to the target variable. However, you have plenty of weakly-labeled instances i.e you have an imperfect version of the target variable for these instances. The idea is that, by modeling the imperfection in the weak labels, we can mitigate the lack of (strongly-labeled) training data. We study three commonly-occurring sources of weak supervision for the land cover mapping problem: (1) Ordinal labels as weak supervision for regression (WORD); (2) Group-level labels as weak supervision for binary classification (WeaSL); and (3) Group-level labels with group-level features as weak supervision for

关键词： Computer science  

摘要： The act of sending unwanted email messages in large quantities to the recipients who have not verifiably given permission to be sent is known as email spam. Even though a lot of times, spam is used as a harmless advertisement, people with malicious intent can take advantage of it and use it to scam people for money or spread computer viruses. Despite the fact that different techniques have been developed to filter spam they are still prevalent, due to the spammer’s rapid adoption of new spam detection techniques. Currently, machine learning techniques are the most effective ways to classify and filter spam emails. The purpose of this study is to conduct a comprehensive study on applying seven popular machine learning techniques (e.g., SVM, ANN, KNN, DT, RF, NB, and LR) to the classification of spam email using the 2007 TREC Public Spam Corpus dataset. Descriptions of the framework of study is given and detail analysis of the study results are provided. The study results provide evidence that the split dataset approach leads to the better performance of classifiers. SVM, RF, KNN, ANN, DT and LR provided 99% accuracy, but NB provided only 92% accuracy. According to other researchers also, NB and RF gave an accuracy of 92% and 99% which is similar to the accuracies of NB and RF classifiers in this study.

关键词： Computer science  

摘要： Visual perception in virtual reality devices is a widely researched topic. Many newer experiments compare their results to those of older studies that may have used equipment which is now outdated, which can cause perceptual differences. These differences in hardware can be simulated to a degree in software, provided the capabilities of the current hardware meet or exceed those of the older hardware. I present the HMD Simulation Framework, a software package for the Unity3D engine that allows for quick modification of many commonly researched HMD characteristics through the Inspector GUI built into Unity. I also describe a human subjects experiment aimed at identifying perceptual equivalence classes between different sets of headset characteristics. Unfortunately, due to the COVID-19 pandemic, all human subjects research was suspended for safety reasons, and I was unable to collect any data.

关键词： Computer architecture   Microprocessors   Training   Clustering algorithms   Decoding   Computer science   RNA   Single cell RNA-seq data   auto-encoder   sparse self-express   spectral clustering  

摘要： Single-cell RNA sequencing(scRNA-seq) technology has boomed in the past decade which makes it possible to study biological problems at the resolution of cellular-level. Currently, the research mainly focuses on exploring the cellular heterogeneity, involving studies about identifying cell type identification, cell lineage tracing, spatial model reconstruction of complex organizations, etc. Clustering analysis is always the most effective way in grouping single cells in previous studies. However, existing scRNA-seq clustering methods separate pre-processing and clustering tasks that complicated the problem. In addition, the emergence of big data further limits the traditional clustering algorithms' application on scRNA-seq data. Therefore, developing novel clustering methods and improving clustering accuracy for growing scRNA-seq data is a continuous task. In this paper, we propose a highly integrated Deep Subspace Clustering Denoise Network named DSCD, which integrates denoise, dimension reduction and clustering in a unified framework. Based on the neural network architecture of autoencoder, DSCD discovers the low dimensional latent structure within scRNA-seq data from the compressed representation. Furthermore, we add a novel self-expressive denoise layer to learning the global relationships between single cells, which is the main innovation of DSCD. Experimental results on the synthetic data demonstrate the effectiveness of the novel denoise layer. From the clustering results on 5 real scRNA-seq datasets, we find that DSCD outperforms the related subspace clustering algorithms and state of the art methods. In conclusion, DSCD responds well to the rapidly increasing scRNA-seq data scale, greatly reduces human interference in dimension reduction and handles the noisy scRNA-seq data in proper way thus obtain a higher clustering accuracy.

关键词： Computer science  

摘要： Theoretical Computer Science has connections to several areas of mathematics, and one of the more prominent of these connections is to combinatorics. Indeed, many problems in this subject are often very combinatorial in nature. These problems have either used existing techniques from combinatorics or have given rise to new combinatorial techniques. This dissertation is a collection of the study of some such problems. 1. We recover a result by Abbe, Shpilka and Wigderson which states that a Reed-Muller code of rate 1 − Θ((logr n)/n) can be recovered from o((log(r−1)/2)/n) randomly chosen errors in a stronger way. Namely, we show that the set of corrupted locations in the message can be recovered just from the syndrome of the message. Among the techniques are the study of tensor decomposition over finite fields and an algorithm to find the roots of a space of low degree polynomials. 2. We show that it is NP-hard to properly 2-color a k-uniform (k − O(√k))-rainbow colorable hypergraph. In particular, we show that it is NP-hard to properly 2-color a 4-uniform 3-rainbow colorable hypergraph. We further extend this using a notion of almost rainbow colorability. We show that given a k-uniform hypergraph where there is a (k − √ck)-coloring of the vertices such that every edge gets (k − 3√ck)-colors, it is NP-hard to properly c-color it. Among the techniques are topological methods to lower bound the chromatic number of certain hypergrpahs, and a theorem of Sarkaria on the chromatic number of the generalized Kneser hypergraph. 3. We show that the discrepancy of a regular hypergraph can be bounded in terms of its spectral information. Let H ⊂ 2[n] be a t-regular hypergraph where |H| ≥ n, and M be the |H| × n incidence matrix. Define λ : = maxv ⊥ 1,||v|| = 1||Mv||. We show that the discrepancy of H is at most O(√t + λ). In particular, this shows that for every t, the discrepancy of a random t-regular hypergraph on m ≥ n hyperedges has discrepancy O(√t) with high probability as

关键词： Computer science  

摘要： Writing declarative models and specifications has numerous benefits, ranging from automatedreasoning and correction of design-level properties before systems are built, to automated testingand debugging of their implementations after they are built. To gain these benefits, the declarativemodel itself needs to be correct. However, correctly writing declarative models that representnon-trivial properties is not easy, especially for practitioners who are not well-versed with theintricate syntax and semantics of declarative *** focus is on Prolog, a declarative programming language rooted in first-order *** supports a robust unit testing framework which allows for developers to systematicallycheck for correctness. However, knowing a model is faulty is just the first step. The developerthen needs to fix the model which means the developer needs to determine what part of themodel is faulty. Prolog models can have numerous interdependent clauses spanning multiplenon-contiguous lines, making the process of locating bugs time consuming. Fault localizationresearch aims to address this problem by automating this process. The key novelty of our workis the introduction of fault localization techniques for Prolog. Our vision is to navigate thecomplex interconnections between Prolog statements for the user; therefore, preventing the userfor analyzing unnecessary portions of their model. However, due to the declarative nature ofProlog, our work explores two different techniques: a spectrum-based technique, which is basedon well-established fault localization techniques for imperative languages, and a mutation-basedtechnique, which is based on a recent advancement in fault *** illustrate theeffectiveness and trade-offs between the two approaches over a set of small but complex Prologmodels.

关键词： Web studies   Educational technology   Computer science  

摘要： Social sharing of photos and videos are one of the most significant uses of the Internet. Open online platforms, like YouTube, are a particularly popular medium for youth who are drawn to the low barrier for entry to reach a potentially global audience. Yet, much of the work that has been carried out focuses on identifying and filtering potentially inappropriate content and otherwise supporting children as consumers of media. This work seeks to understand and support youth as video creators. In this thesis, I will describe my foundational work investigating youth video creation on commercial platforms, present a novel exploration of design opportunities with a team of intergenerational youth and detail one specific design insight that blossomed from this exploration, the opportunity for collaboration in video creation.