Thesis (M.S.)--Oregon State University, 2004. / Printout. Includes bibliographical references (leaves 45-46). Also available on the World Wide Web.

To perform close proximity operations under a low-gravity environment, relative and absolute positions are vital information to the maneuver. Hence navigation is inseparably integrated in space travel. Extreme Learning Machine (ELM) is presented as an optical navigation method around small celestial bodies. Optical Navigation uses visual observation instruments such as a camera to acquire useful data and determine spacecraft position. The required input data for operation is merely a single image strip and a nadir image. ELM is a machine learning Single Layer feed-Forward Network (SLFN), a type of neural network (NN). The algorithm is developed on the predicate that input weights and biases can be randomly assigned and does not require back-propagation. The learned model is the output layer weights which are used to calculate a prediction. Together, Extreme Learning Machine Optical Navigation (ELM OpNav) utilizes optical images and ELM algorithm to train the machine to navigate around a target body. In this thesis the asteroid, Vesta, is the designated celestial body. The trained ELMs estimate the position of the spacecraft during operation with a single data set. The results show the approach is promising and potentially suitable for on-board navigation.

The rise of deep learning has facilitated remarkable progress in video understanding. This thesis addresses three important tasks of video understanding: video object detection, joint object and action detection, and spatio-temporal action localization. Object class detection is one of the most important challenges in computer vision. Object detectors are usually trained on bounding-boxes from still images. Recently, video has been used as an alternative source of data. Yet, training an object detector on one domain (either still images or videos) and testing on the other one results in a significant performance gap compared to training and testing on the same domain. In the first part of this thesis, we examine the reasons behind this performance gap. We define and evaluate several domain shift factors: spatial location accuracy, appearance diversity, image quality, aspect distribution, and object size and camera framing. We examine the impact of these factors by comparing the detection performance before and after cancelling them out. The results show that all five factors affect the performance of the detectors and their combined effect explains the performance gap. While most existing approaches for detection in videos focus on objects or human actions separately, in the second part of this thesis we aim at detecting non-human centric actions, i.e., objects performing actions, such as cat eating or dog jumping. We introduce an end-to-end multitask objective that jointly learns object-action relationships. We compare it with different training objectives, validate its effectiveness for detecting object-action pairs in videos, and show that both tasks of object and action detection benefit from this joint learning. In experiments on the A2D dataset [Xu et al., 2015], we obtain state-of-the-art results on segmentation of object-action pairs. In the third part, we are the first to propose an action tubelet detector that leverages the temporal continuity of videos instead of operating at the frame level, as state-of-the-art approaches do. The same way modern detectors rely on anchor boxes, our tubelet detector is based on anchor cuboids by taking as input a sequence of frames and outputing tubelets, i.e., sequences of bounding boxes with associated scores. Our tubelet detector outperforms all state of the art on the UCF-Sports [Rodriguez et al., 2008], J-HMDB [Jhuang et al., 2013a], and UCF-101 [Soomro et al., 2012] action localization datasets especially at high overlap thresholds. The improvement in detection performance is explained by both more accurate scores and more precise localization.

Made available in DSpace on 2014-12-17T14:56:02Z (GMT). No. of bitstreams: 1 MeikaIM.pdf: 766418 bytes, checksum: 87a604688aa5cd2c4f6aba8237c67210 (MD5) Previous issue date: 2005-12-13 / One of the most important goals of bioinformatics is the ability to identify genes in uncharacterized DNA sequences on world wide database. Gene expression on prokaryotes initiates when the RNA-polymerase enzyme interacts with DNA regions called promoters. In these regions are located the main regulatory elements of the transcription process. Despite the improvement of in vitro techniques for molecular biology analysis, characterizing and identifying a great number of promoters on a genome is a complex task. Nevertheless, the main drawback is the absence of a large set of promoters to identify conserved patterns among the species. Hence, a in silico method to predict them on any species is a challenge. Improved promoter prediction methods can be one step towards developing more reliable ab initio gene prediction methods. In this work, we present an empirical comparison of Machine Learning (ML) techniques such as Na??ve Bayes, Decision Trees, Support Vector Machines and Neural Networks, Voted Perceptron, PART, k-NN and and ensemble approaches (Bagging and Boosting) to the task of predicting Bacillus subtilis. In order to do so, we first built two data set of promoter and nonpromoter sequences for B. subtilis and a hybrid one. In order to evaluate of ML methods a cross-validation procedure is applied. Good results were obtained with methods of ML like SVM and Na?ve Bayes using B. subtilis. However, we have not reached good results on hybrid database / Um dos grandes desafios da Bioinform?tica ? manipular e analisar os dados acumulados nas bases de dados mundiais. A express?o dos genes em procariotos ? iniciada quando a enzima RNA polimerase une-se com uma regi?o pr?xima ao gene, chamada de promotor, onde ? localizado os principais elementos regulat?rios do processo de transcri??o. Apesar do crescente avan?o das t?cnicas experimentais (in vitro) em biologia molecular, caracterizar e identificar um n?mero significante de promotores ainda ? uma tarefa dif?cil. Os m?todos computacionais existentes enfrentam a falta de um n?mero adequado de promotores conhecidos para identificar padr?es conservados entre as esp?cies. Logo, um m?todo para prediz?-los em qualquer organismo procari?tico ainda ? um desafio. Neste trabalho, apresentamos uma compara??o emp?rica de t?cnicas individuais de aprendizado de m?quina, tais como: Classificador Bayesiano Ing?nuo, ?rvores de Decis?o, M?quinas de Vetores de Suporte, Redes Neurais do tipo VotedPerceptron, PART e k-Vizinhos Mais Pr?ximos e sistemas multiclassificadores (Bagging e Adaboosting) e Modelo Oculto de Markov ? tarefa de predi??o de promotores procariotos em Bacilos subtilis. Utilizamos a valida??o cruzada para avaliar todos os m?todos de AM. Para esses testes, foram constru?das base de dados com seq??ncias de promotores e n?o-promotores do Bacillus subtilis e uma base de dados h?brida. Os m?todos de AM obtiveram bons resultados com o SVM e o Na?ve Bayes. N?o conseguimos entretanto, obter resultados relevantes para a base de dados h?brida

Nesta tese, aborda-se o problema de classificaÃÃo de dados que estÃo contaminados com pa- drÃes atÃpicos. Tais padrÃes, genericamente chamados de outliers, sÃo onipresentes em conjunto de dados multivariados reais, porÃm sua detecÃÃo a priori (i.e antes de treinar um classificador) à uma tarefa de difÃcil realizaÃÃo. Como conseqÃÃncia, uma abordagem reativa, em que se desconfia da presenÃa de outliers somente apÃs um classificador previamente treinado apresen- tar baixo desempenho, à a mais comum. VÃrias estratÃgias podem entÃo ser levadas a cabo a fim de melhorar o desempenho do classificador, dentre elas escolher um classificador mais poderoso computacionalmente ou promover uma limpeza dos dados, eliminando aqueles pa- drÃes difÃceis de categorizar corretamente. Qualquer que seja a estratÃgia adotada, a presenÃa de outliers sempre irà requerer maior atenÃÃo e cuidado durante o projeto de um classificador de padrÃes. Tendo estas dificuldades em mente, nesta tese sÃo revisitados conceitos e tÃcni- cas provenientes da teoria de regressÃo robusta, em particular aqueles relacionados à estimaÃÃo M, adaptando-os ao projeto de classificadores de padrÃes capazes de lidar automaticamente com outliers. Esta adaptaÃÃo leva à proposiÃÃo de versÃes robustas de dois classificadores de padrÃes amplamente utilizados na literatura, a saber, o classificador linear dos mÃnimos qua- drados (least squares classifier, LSC) e a mÃquina de aprendizado extremo (extreme learning machine, ELM). AtravÃs de uma ampla gama de experimentos computacionais, usando dados sintÃticos e reais, mostra-se que as versÃes robustas dos classificadores supracitados apresentam desempenho consistentemente superior aos das versÃes originais. / This thesis addresses the problem of data classification when they are contaminated with atypical patterns. These patterns, generally called outliers, are omnipresent in real-world multi- variate data sets, but their a priori detection (i.e. before training the classifier) is a difficult task to perform. As a result, the most common approach is the reactive one, in which one suspects of the presence of outliers in the data only after a previously trained classifier has achieved a low performance. Several strategies can then be carried out to improve the performance of the classifier, such as to choose a more computationally powerful classifier and/or to remove the de- tected outliers from data, eliminating those patterns which are difficult to categorize properly. Whatever the strategy adopted, the presence of outliers will always require more attention and care during the design of a pattern classifier. Bearing these difficulties in mind, this thesis revi- sits concepts and techniques from the theory of robust regression, in particular those related to M-estimation, adapting them to the design of pattern classifiers which are able to automatically handle outliers. This adaptation leads to the proposal of robust versions of two pattern classi- fiers widely used in the literature, namely, least squares classifier (LSC) and extreme learning machine (ELM). Through a comprehensive set of computer experiments using synthetic and real-world data, it is shown that the proposed robust classifiers consistently outperform their original versions.

Positive natural selection leaves detectable, distinctive patterns in the genome in the form of a selective sweep. Identifying areas of the genome that have undergone selective sweeps is an area of high interest as it enables understanding of species and population evolution. Previous work has accomplished this by evaluating patterns within summary statistics computed across the genome and through application of machine learning techniques to raw population genomic data. When using raw population genomic data, convolutional neural networks have most recently been employed as they can handle large input arrays and maintain correlations among elements. Yet, such models often require massive amounts of training data and can be computationally expensive to train for a given problem. Instead, transfer learning has recently been used in the image analysis literature to improve machine learning models by learning the important features of images from large unrelated datasets beforehand, and then refining these models through subsequent application on smaller and more relevant datasets. We combine transfer learning with convolutional neural networks to improve classification of selective sweeps from raw population genomic data. We show that the combination of transfer learning with convolutional neural networks allows for accurate classification of selective sweeps. / Includes bibliography. / Thesis (M.S.)--Florida Atlantic University, 2021. / FAU Electronic Theses and Dissertations Collection

Dans de nombreux problèmes en apprentissage supervisé, les entrées ont une structure de groupes connue et/ou clairement identifiable. Dans ce contexte, l'élaboration d'une règle de prédiction utilisant les groupes plutôt que les variables individuelles peut être plus pertinente tant au niveau des performances prédictives que de l'interprétation. L'objectif de la thèse est de développer des méthodes par arbres adaptées aux variables groupées. Nous proposons deux approches qui utilisent la structure groupée des variables pour construire des arbres de décisions. La première méthode permet de construire des arbres binaires en classification. Une coupure est définie par le choix d'un groupe et d'une combinaison linéaire des variables du dit groupe. La seconde approche, qui peut être utilisée en régression et en classification, construit un arbre non-binaire dans lequel chaque coupure est un arbre binaire. Ces deux approches construisent un arbre maximal qui est ensuite élagué. Nous proposons pour cela deux stratégies d'élagage dont une est une généralisation du minimal cost-complexity pruning. Les arbres de décision étant instables, nous introduisons une méthode de forêts aléatoires pour variables groupées. Outre l'aspect prédiction, ces méthodes peuvent aussi être utilisées pour faire de la sélection de groupes grâce à l'introduction d'indices d'importance des groupes. Ce travail est complété par une partie indépendante dans laquelle nous nous plaçons dans un cadre d'apprentissage non supervisé. Nous introduisons un nouvel algorithme de clustering. Sous des hypothèses classiques, nous obtenons des vitesses de convergence pour le risque de clustering de l'algorithme proposé. / In many problems in supervised learning, inputs have a known and/or obvious group structure. In this context, elaborating a prediction rule that takes into account the group structure can be more relevant than using an approach based only on the individual variables for both prediction accuracy and interpretation. The goal of this thesis is to develop some tree-based methods adapted to grouped variables. Here, we propose two new tree-based approaches which use the group structure to build decision trees. The first approach allows to build binary decision trees for classification problems. A split of a node is defined according to the choice of both a splitting group and a linear combination of the inputs belonging to the splitting group. The second method, which can be used for prediction problems in both regression and classification, builds a non-binary tree in which each split is a binary tree. These two approaches build a maximal tree which is next pruned. To this end, we propose two pruning strategies, one of which is a generalization of the minimal cost-complexity pruning algorithm. Since decisions trees are known to be unstable, we introduce a method of random forests that deals with groups of inputs. In addition to the prediction purpose, these new methods can be also use to perform group variable selection thanks to the introduction of some measures of group importance, This thesis work is supplemented by an independent part in which we consider the unsupervised framework. We introduce a new clustering algorithm. Under some classical regularity and sparsity assumptions, we obtain the rate of convergence of the clustering risk for the proposed alqorithm.

The rapid growth of data-intensive workloads such as deep learning and artificial intelligence has placed significant strain on the interconnects of high performance computing systems, presenting a looming bottleneck of significant societal concern. Furthermore, with the impending end of Moore's Law, continued reliance on transistor density scaling in compute nodes to compensate for this bottleneck will experience an abrupt halt in the coming decade. Optical interconnects provide an appealing path to mitigating this communication bottleneck through leveraging the favorable physical properties of light to increase bandwidth while simultaneously reducing energy consumption with distance-agnostic performance, in stark contrast to electrical signaling. In particular, silicon photonics presents an ideal platform for optical interconnects for a variety of economic, fundamental scientific, and engineering reasons; namely, (i) the chips are fabricated using the same mature complementary metal-oxide-semiconductor (CMOS) infrastructure used for microelectronic chips; (ii) the high index contrast between silicon and silicon dioxide permits micron-scale devices at telecommunication wavelengths; and (iii) decades of engineering effort has resulted in state-of-the-art devices comparable to discrete components in other material platforms including low-loss (< 0.5 dB/cm) waveguides, high-speed (> 100 Gb/s) modulators and photodetectors, and low-loss (< 1 dB) fiber-to-chip interfaces. Through leveraging these favorable properties of the platform, silicon photonic chips can be directly co-packaged with CMOS electronics to yield unprecedented interconnect bandwidth at length scales ranging from millimeters to kilometers while simultaneously achieving substantial reduction in energy consumption relative to currently deployed solutions. The work in this thesis aims to address the fundamental scalability of silicon photonic interconnects to orders-of-magnitude beyond the current state-of-the-art, enabling extreme channel counts in the frequency domain through leveraging advances in chip-scale Kerr frequency combs. While the current co-packaged optics roadmap includes silicon photonics as an enabling technology to ~ 5 pJ/bit terabit-scale interconnects, this work examines the foundational challenges which must be overcome to realize forward-looking sub-pJ/bit petabit-scale optical I/O. First, an overview of the system-level challenges associated with such links is presented, motivating the following chapters focused on device innovations that address these challenges. Leveraging these advances, a novel link architecture capable of scaling to hundreds of wavelength channels is proposed and experimentally demonstrated, providing an appealing path to future petabit/s photonic interconnects with sub-pJ/bit energy consumption. Such photonic interconnects with ultra-high bandwidth, ultra-low energy consumption, and low latency have the potential to revolutionize future data center and high performance computing systems through removing the strong constraint of data locality, permitting drastically new architectures through resource disaggregation. The advances demonstrated in this thesis provide a clear direction towards realizing future green hyper-scale data centers and high performance computers with environmentally-conscious scaling, providing an energy-efficient and massively scalable platform capable of keeping pace with ever-growing bandwidth demands through the next quarter-century and beyond.

In this dissertation, we are concerned with the advancement of optimization algorithms for training deep learning models, and in particular about practical second-order methods that take into account the structure of deep neural networks (DNNs). Although first-order methods such as stochastic gradient descent have long been the predominant optimization algorithm used in deep learning, second-order methods are of interest because of their ability to use curvature information to accelerate the optimization process. After the presentation of some background information in Chapter 1, Chapters 2 and 3 focus on the development of practical quasi-Newton methods for training DNNs. We analyze the Kronecker-factored structure of the Hessian matrix of multi-layer perceptrons and convolutional neural networks and consequently propose block-diagonal Kronecker-factored quasi-Newton methods named K-BFGS and K-BFGS(L). To handle the non-convexity nature of DNNs, we also establish new double damping techniques for our proposed methods. Our K-BFGS and K-BFGS(L) methods have memory requirements comparable to first-order methods and experience only mild overhead in terms of per-iteration time complexity. In Chapter 4, we develop a new approximate natural gradient method named Tensor Normal Training (TNT), in which the Fisher matrix is viewed as the covariance matrix of a tensor normal distribution (a generalized form of the normal distribution). The tractable Kronecker-factored approximation to the Fisher information matrix that results from this approximation enables TNT to enjoy memory requirements and per-iteration computational costs that are only slightly higher than those for first-order methods. Notably, unlike KFAC and K-BFGS/K-BFGS(L), TNT only requires the knowledge of the shape of the trainable parameters of a model and does not depend on the specific model architecture. In Chapter 5, we consider the subsampled versions of Gauss-Newton and natural gradient methods applied to DNNs. Because of the low-rank nature of the subsampled matrices, we make use of the Sherman-Morrison-Woodbury formula along with backpropagation to efficiently compute their inverse. We also show that, under rather mild conditions, the algorithm converges to a stationary point if Levenberg-Marquardt damping is used. The results of a substantial number of numerical experiments are reported in Chapters 2, 3, 4 and 5, in which we compare the performance of our methods to state-of-the-art methods used to train DNNs, that demonstrate the efficiency and effectiveness of our proposed new second-order methods.

Testing analog circuits is more difficult than digital circuits. The reasons for this difficulty include continuous time and amplitude signals, lack of well-accepted testing techniques and time and cost required for its realization. The traditional method for testing analog circuits involves measuring all the performance parameters and comparing the measured parameters with the limits of the data-sheet specifications. Because of the large number of data-sheet specifications, the test generation and application requires long test times and expensive test equipment. This thesis proposes an implicit functional testing technique for analog circuits that can be easily implemented in BIST circuitry. The proposed technique does not require measuring data-sheet performance parameters. To simplify the testing only time domain digital input is required. For each circuit under test (CUT) a cross-covariance signature is computed from the test input and CUT's output. The proposed method requires a training sample of the CUT to be binned to the data-sheet specifications. The binned CUT sample cross-covariance signatures are mapped with a supervised machine learning classifier. For each bin, the classifiers select unique sub-sets of the cross-covariance signature. The trained classifier is then used to bin newly manufactured copies of the CUT. The proposed technique is evaluated on synthetic data generated from the Monte Carlo simulation of the nominal circuit. Results show the machine learning classifier must be chosen to match the imbalanced bin populations common in analog circuit testing. For sample sizes of 700+ and training for individual bins, classifier test escape rates ranged from 1000 DPM to 10,000 DPM.