教学课程资源库 更多>>

关键词： Computer science   Artificial intelligence   Computer engineering  

摘要： Computer Vision has benefited significantly from Deep Learning architectures, such as Convolutional Neural Networks (CNNs). Semantic Segmentation, which is pixel-wise image classification, is a key area of research within Computer Vision that has seen much growth from CNNs as well. With stand-alone embedded computer hardware becoming more capable and prominent, more applications of Semantic Segmentation have arisen. Thus, current technologies have been augmented, namely autonomous vehicles, and new technologies, including augmented reality (AR) and virtual reality (VR), have emerged in recent years. The effectiveness and performance of these technologies are dependent on object detection and, as a result, Semantic Segmentation has become a key aspect of research and development. CNNs designed for Semantic Segmentation tasks have shown significant performance in scene understanding, however most CNNs for these tasks require copious amounts of data as well as extensive computational resources. Consequently, current methods using CNNs for Semantic Segmentation tasks lack the real-time processing capabilities required for these technologies that are employed on embedded systems. In addition, the amount of annotated data readily available for AR/VR technologies is limited, rendering conventional training methods ineffective. In this work, a Minimized Efficient Network (MinENet) architecture is first shown to improve upon accuracy of Semantic Segmentation designed for an embedded AR/VR system. Second, this research presents EyeSeg, an encoder-decoder architecture, designed for accurate Semantic Segmentation with sparsely annotated data and applied to similar AR/VR problems. Lastly, CitySeg, an encoder-decoder architecture augmented with convolutional long-short term memory units (ConvLSTM), is presented as an extension of this work. CitySeg's preliminary results showcase the capabilities of efficient lightweight architectures while maintaining accuracy on data that has si

关键词： Game theory   Decomposition   Convex analysis   Partial differential equations   Algorithms   Equilibrium   System theory   Computer science   Mathematics  

摘要： This dissertation develops novel nonlinear analysis tools and methodologies to advance the field of monotone operator theory and its applications. First, we show that the Douglas-Rachford algorithm and Spingarn’s method of partial inverses can fail to converge strongly. Second, we introduce the notion of a warped resolvent as an extension of the classical resolvent, study its properties, and propose weakly and strongly convergent warped proximal iteration principles. This framework unifies and extends several so-far unrelated ones, such as projective splitting and Tseng’s forward-backward-forward method. In addition, the warped resolvent framework is shown to be an effective device to produce new and flexible splitting methods for complex monotone inclusion problems. Next, we introduce a Bregman forward-backward method for solving monotone inclusions, establish its weak convergence, and show that it captures and extends several iterative methods. Another contribution is to propose a new saddle presentation to study and solve highly structured systems of monotone inclusions. This leads to highly flexible asynchronous block-iterative algorithms. Applications of warped resolvents in the context of variational inequalities, convex optimization, Nash equilibrium, and network flows are discussed. Finally, we analyze the merits and performance of block-activated algorithms for solving multicomponent fully nonsmooth minimization problems with applications to machine learning and image recovery.

关键词： Lighting   Estimation   Image color analysis   Visualization   Image enhancement   Distortion   Computer science   Underexposed photo enhancement   perceptually bidirectional similarity   illumination estimation  

摘要： Although remarkable progress has been made, existing methods for enhancing underexposed photos tend to produce visually unpleasing results due to the existence of visual artifacts (e.g., color distortion, loss of details and uneven exposure). We observed that this is because they fail to ensure the perceptual consistency of visual information between the source underexposed image and its enhanced output. To obtain high-quality results free of these artifacts, we present a novel underexposed photo enhancement approach that is able to maintain the perceptual consistency. We achieve this by proposing an effective criterion, referred to as perceptually bidirectional similarity, which explicitly describes how to ensure the perceptual consistency. Particularly, we adopt the Retinex theory and cast the enhancement problem as a constrained illumination estimation optimization, where we formulate perceptually bidirectional similarity as constraints on illumination and solve for the illumination which can recover the desired artifact-free enhancement results. In addition, we describe a video enhancement framework that adopts the presented illumination estimation for handling underexposed videos. To this end, a probabilistic approach is introduced to propagate illuminations of sampled keyframes to the entire video by tackling a Bayesian Maximum A Posteriori problem. Extensive experiments demonstrate the superiority of our method over the state-of-the-art methods.

关键词： Bioinformatics   Genetics   Computer science  

摘要： The era of personalized genetic medicine has arrived. It is now routine to sequence an individual’s DNA, either whole genome sequences (WGS) or targeted, as part of a patient’s medical plan. One of the exceptions is an ~200 kilobase region in chromosome 19 containing the killer-cell immunoglobulin-like receptors (KIR). These genes encode proteins that influence the actions of natural killer (NK) cells based on whether or not they bind with peptide-bound human leukocyte antigen receptors. This is evolutionary important to fight pathogens and mediate pregnancy. Modern medicine has correlated low-resolution KIR with many diseases and treatments, although the findings are often relatively vague and sometimes contradictory due to low resolution interpretation and/or small cohort sizes. Whether for personal medicine or population studies, the current best practices for KIR genotyping are to determine the presence/absence (PA) or copy number variation (CNV) of each gene using oligo- or primer-based polymerase chain reaction. The goal of our research is to advance DNA sequencing and interpretation of human KIR haplotypes. To that end, we have created algorithms and workflows to enhance interpretations of KIR at resolutions from PA genotyping via short-read WGS to full-haplotype assembly from long-read targeted or whole-genome sequences. First, we developed the first workflow to efficiently and accurately capture, sequence, assemble, and annotate full KIR haplotype sequences. As part of this workflow, we designed small sequences to capture the DNA fragments. Next, we use the alignment pattern of those short sequences across finished KIR haplotypes to define and annotate haplotype structures. The results show, for the first time, that the KIR region is composed of 9 genes in 14 loci. Next, we annotated all 68 reported human haplotypes, aligned them at the structural level, and then refined the alignment down the base level, providing the first KIR haplotype multiple sequence

关键词： Algebra   Quaternions   Rotors   Three-dimensional displays   Robots   Computer science   Calculus   Geometric algebra   Clifford algebra   quaternion algebra   screw theory   signal processing   electrical engineering and power systems   geometric and quantum computing   image processing   computer vision   graphic engineering   artificial intelligence   machine learning   neural networks   control engineering   robotics   biomedical engineering   and biotechnology  

摘要： Geometric Algebra (GA) has proven to be an advanced language for mathematics, physics, computer science, and engineering. This review presents a comprehensive study of works on Quaternion Algebra and GA applications in computer science and engineering from 1995 to 2020. After a brief introduction of GA, the applications of GA are reviewed across many fields. We discuss the characteristics of the applications of GA to various problems of computer science and engineering. In addition, the challenges and prospects of various applications proposed by many researchers are analyzed. We analyze the developments using GA in image processing, computer vision, neurocomputing, quantum computing, robot modeling, control, and tracking, as well as improvement of computer hardware performance. We believe that up to now GA has proven to be a powerful geometric language for a variety of applications. Furthermore, there is evidence that this is the appropriate geometric language to tackle a variety of existing problems and that consequently, step-by-step GA-based algorithms should continue to be further developed. We also believe that this extensive review will guide and encourage researchers to continue the advancement of geometric computing for intelligent machines.

关键词： Civil engineering   Mechanics   Computer science   Environmental geology  

关键词： Computer science   Computer engineering   Artificial intelligence   Information technology  

摘要： The rise of deep learning and advance neural network architectures has seen a rise of conversational agents that have near human text generation capabilities. However, these models, which are trained using supervised learning are constrained by their objective function which is the Maximum Likelihood Estimate. Reinforcement learning is a good alternative to supervised learning but it faces some limitations. The primary limitation is the lack of a definition for a good reward signal to train such conversational agents. In our thesis we provide our solution to this problem and compare the models trained using our reward to other baselines. Finally we compare the strengths and weakness of our reward signal by testing the robustness of the model to corruption schemes and its ability to explore diverse responses.

关键词： Computer science   Physiology   Artificial intelligence  

摘要： Mind wandering (MW) is a phenomenon where a person shifts their attention from task-related to task-unrelated information. Mind Wandering is an omnipresent phenomenon for human beings. The consequence of mind wandering can impact a person’s performance negatively. Reorienting the mind’s attention using technology shows great promise to improve the performance and productivity of people in learning or other performative tasks. In this research, we investigate 62 eye gaze features by dividing them into four sets of global features: eye movement descriptive features, pupil diameter descriptive features, blink features, and miscellaneous features to detect mind wandering during reading from a computer interface. Our dataset, which was collected from a previous study, contains a mind wandering report where 135 participants were recorded “mind wandering” or “not mind wandering” using self-reporting during a computerized reading task. During this process, a remotely placed eye tracker tool recorded eye gaze data. Models were created using six supervised conventional machine learning (ML) algorithms: logistic regression, k-nearest neighbors (k-NN), support vector machine (SVM), decision tree, random forest and naive Bayes. Machine learning models were trained on eye gaze dataset and evaluated using 5-fold cross validation. We measured the performance using area under the receiver operating characteristics (AUC-ROC) score, AUC-ROC curve, and confusion-matrix. To further improve the AUC-ROC score and other evaluation metrics, we trained standard neural networks and deep learning models using the data. Four sets of deep learning architectures were trained and evaluated. We found that dense neural network with one dimensional convolutional layer (DNN+Conv1D) outperformed the performance of conventional machine learning models. Naïve Bayes achieved mean test AUC-ROC score of 0.6595 and mean test accuracy of 0.6416. DNN+Conv1D beat the AUC-ROC score and achieved a score of 0.8024

关键词： Computer science   Computer engineering   Applied mathematics   Artificial intelligence  

摘要： This dissertation largely studies problems of two types. In the first part, we study ranking and clustering in the presence of outliers. The goal is to remove up to a specific number of elements and then find a more meaningful clustering/ranking on the remaining ones. - Rank Aggregation with Outliers. In the rank aggregation from pairwise comparisons, the objective is to order n given elements by aggregating pairwise comparison information which could be inconsistent. Inconsistencies caused by outliers can highly affect the final ranking. We initiate the study of rank aggregation with outliers in which we are allowed to discard a small fraction of the elements as outliers to get a better ranking. - k-means Clustering with Outliers. In the standard k-means clustering we are given a set of n points in Euclidean space (or any metric space) and a target number of clusters k and the goal is to choose a set of k points from the space as centers, so as to minimize the sum of the squared distances of every point to its closest center. Since k-means clustering is highly sensitive to noise, we study it in the presence of outliers by allowing to discard a small fraction of points. The previous algorithms for k-means clustering with outliers are either complicated, do not have provable guarantees, suffer from high running times or are not applicable in the distributed settings. The algorithms we develop address these challenges simultaneously. In the second part of this dissertation, we bring our work on augmenting machine learned prediction to traditional algorithm design for online problems, specifically for extensions of two classical online problems, non-clairvoyant job scheduling and knapsack. Algorithms developed in this model benefit from the prediction given by the estimator. The goal is to get better results if the prediction is accurate and get results not much worse than traditional algorithms guarantee even with error-prone predictions. - Non-Clairvoyant Scheduling

关键词： Metals   Polymers   Acids   Hydrocarbons   Carbon   Equilibrium   Hydrogen   Alcohol   Energy   Algorithms   Clustering   Geometry   Morphology   Computer science   Mathematics   Polymer chemistry  

摘要： Despite the ubiquitous use of polymers in every technological arena, a surprising amount of theoretical understanding has yet to be verified due to the notorious difficulty of their computational modeling. This work addresses this issue on multiple fronts, with multiple length scales and combinations among them. On the smallest length scale, we compared two assumed esterification mechanism using density functional theory (DFT), validating one reaction network showing that prevalent ligand exchange and proton donation contribute to an intramolecular acyl transfer. Using coarse-grained (CG) dissipative particle dynamics (DPD) we modeled proton diffusion in a polyelectrolyte membrane, demonstrating and characterizing nanophase separation and the resulting influence on transport properties. As an alternative to DPD, we sought to develop a CG model with more chemical specificity. To achieve this, we used the SAFT--Mie formalism which explicitly connects nonbonded potentials of molecular segments to thermodynamic properties through an equation of state. Leveraging experience with DFT, we derived a method to calculate cross-interaction parameters from multipole moments obtained using DFT, in a completely predictive manner. Then, to make this complex process easy to implement, we launched the distributed python packages, MAPSCI and DESPASITO. Through this work, we have improved mechanistic understanding through both electronic structure methods and coarse-grained simulations and built the ability for continued and more chemically specific study polymeric research.