New Techniques in Deep Representation Learning

Thesis (Ph.D.)--University of Washington, 2016-06

Applying machine learning to catalogue matching in astrophysics

We present the results of applying automated machine learning techniques to the problem of matching different object catalogues in astrophysics. In this study, we take two partially matched catalogues where one of the two catalogues has a large positional uncertainty. The two catalogues we used here were taken from the H I Parkes All Sky Survey (HIPASS) and SuperCOSMOS optical survey. Previous work had matched 44 per cent (1887 objects) of HIPASS to the SuperCOSMOS catalogue. A supervised learning algorithm was then applied to construct a model of the matched portion of our catalogue. Validation of the model shows that we achieved a good classification performance (99.12 per cent correct). Applying this model to the unmatched portion of the catalogue found 1209 new matches. This increases the catalogue size from 1887 matched objects to 3096. The combination of these procedures yields a catalogue that is 72 per cent matched.

An Incremental EM-based Learning Approach for On-Line Prediction of Hospital Resource Utilization

Objective: Inpatient length of stay (LOS) is an important measure of hospital activity, health care resource consumption, and patient acuity. This research work aims at developing an incremental expectation maximization (EM) based learning approach on mixture of experts (ME) system for on-line prediction of LOS. The use of a batchmode learning process in most existing artificial neural networks to predict LOS is unrealistic, as the data become available over time and their pattern change dynamically. In contrast, an on-line process is capable of providing an output whenever a new datum becomes available. This on-the-spot information is therefore more useful and practical for making decisions, especially when one deals with a tremendous amount of data. Methods and material: The proposed approach is illustrated using a real example of gastroenteritis LOS data. The data set was extracted from a retrospective cohort study on all infants born in 1995-1997 and their subsequent admissions for gastroenteritis. The total number of admissions in this data set was n = 692. Linked hospitalization records of the cohort were retrieved retrospectively to derive the outcome measure, patient demographics, and associated co-morbidities information. A comparative study of the incremental learning and the batch-mode learning algorithms is considered. The performances of the learning algorithms are compared based on the mean absolute difference (MAD) between the predictions and the actual LOS, and the proportion of predictions with MAD < 1 day (Prop(MAD < 1)). The significance of the comparison is assessed through a regression analysis. Results: The incremental learning algorithm provides better on-line prediction of LOS when the system has gained sufficient training from more examples (MAD = 1.77 days and Prop(MAD < 1) = 54.3%), compared to that using the batch-mode learning. The regression analysis indicates a significant decrease of MAD (p-value = 0.063) and a significant (p-value = 0.044) increase of Prop(MAD

The exact sample complexity of PAC-learning problems with unit VC dimension

The Vapnik-Chervonenkis (VC) dimension is a combinatorial measure of a certain class of machine learning problems, which may be used to obtain upper and lower bounds on the number of training examples needed to learn to prescribed levels of accuracy. Most of the known bounds apply to the Probably Approximately Correct (PAC) framework, which is the framework within which we work in this paper. For a learning problem with some known VC dimension, much is known about the order of growth of the sample-size requirement of the problem, as a function of the PAC parameters. The exact value of sample-size requirement is however less well-known, and depends heavily on the particular learning algorithm being used. This is a major obstacle to the practical application of the VC dimension. Hence it is important to know exactly how the sample-size requirement depends on VC dimension, and with that in mind, we describe a general algorithm for learning problems having VC dimension 1. Its sample-size requirement is minimal (as a function of the PAC parameters), and turns out to be the same for all non-trivial learning problems having VC dimension 1. While the method used cannot be naively generalised to higher VC dimension, it suggests that optimal algorithm-dependent bounds may improve substantially on current upper bounds.

Incorporating curvature information into on-line learning

We analyse the dynamics of a number of second order on-line learning algorithms training multi-layer neural networks, using the methods of statistical mechanics. We first consider on-line Newton's method, which is known to provide optimal asymptotic performance. We determine the asymptotic generalization error decay for a soft committee machine, which is shown to compare favourably with the result for standard gradient descent. Matrix momentum provides a practical approximation to this method by allowing an efficient inversion of the Hessian. We consider an idealized matrix momentum algorithm which requires access to the Hessian and find close correspondence with the dynamics of on-line Newton's method. In practice, the Hessian will not be known on-line and we therefore consider matrix momentum using a single example approximation to the Hessian. In this case good asymptotic performance may still be achieved, but the algorithm is now sensitive to parameter choice because of noise in the Hessian estimate. On-line Newton's method is not appropriate during the transient learning phase, since a suboptimal unstable fixed point of the gradient descent dynamics becomes stable for this algorithm. A principled alternative is to use Amari's natural gradient learning algorithm and we show how this method provides a significant reduction in learning time when compared to gradient descent, while retaining the asymptotic performance of on-line Newton's method.

Protein-protein interactions classification from text via local learning with class priors

Text classification is essential for narrowing down the number of documents relevant to a particular topic for further pursual, especially when searching through large biomedical databases. Protein-protein interactions are an example of such a topic with databases being devoted specifically to them. This paper proposed a semi-supervised learning algorithm via local learning with class priors (LL-CP) for biomedical text classification where unlabeled data points are classified in a vector space based on their proximity to labeled nodes. The algorithm has been evaluated on a corpus of biomedical documents to identify abstracts containing information about protein-protein interactions with promising results. Experimental results show that LL-CP outperforms the traditional semisupervised learning algorithms such as SVMand it also performs better than local learning without incorporating class priors.

The evolution of learning systems:to Bayes or not to be

Bayesian algorithms pose a limit to the performance learning algorithms can achieve. Natural selection should guide the evolution of information processing systems towards those limits. What can we learn from this evolution and what properties do the intermediate stages have? While this question is too general to permit any answer, progress can be made by restricting the class of information processing systems under study. We present analytical and numerical results for the evolution of on-line algorithms for learning from examples for neural network classifiers, which might include or not a hidden layer. The analytical results are obtained by solving a variational problem to determine the learning algorithm that leads to maximum generalization ability. Simulations using evolutionary programming, for programs that implement learning algorithms, confirm and expand the results. The principal result is not just that the evolution is towards a Bayesian limit. Indeed it is essentially reached. In addition we find that evolution is driven by the discovery of useful structures or combinations of variables and operators. In different runs the temporal order of the discovery of such combinations is unique. The main result is that combinations that signal the surprise brought by an example arise always before combinations that serve to gauge the performance of the learning algorithm. This latter structures can be used to implement annealing schedules. The temporal ordering can be understood analytically as well by doing the functional optimization in restricted functional spaces. We also show that there is data suggesting that the appearance of these traits also follows the same temporal ordering in biological systems. © 2006 American Institute of Physics.

Local Goals Driven Hierarchical Reinforcement Learning

* This research was partially supported by the Latvian Science Foundation under grant No.02-86d.

A Comparative Analysis of Predictive Learning Algorithms on High-Dimensional Microarray Cancer Data

This research evaluates pattern recognition techniques on a subclass of big data where the dimensionality of the input space (p) is much larger than the number of observations (n). Specifically, we evaluate massive gene expression microarray cancer data where the ratio κ is less than one. We explore the statistical and computational challenges inherent in these high dimensional low sample size (HDLSS) problems and present statistical machine learning methods used to tackle and circumvent these difficulties. Regularization and kernel algorithms were explored in this research using seven datasets where κ < 1. These techniques require special attention to tuning necessitating several extensions of cross-validation to be investigated to support better predictive performance. While no single algorithm was universally the best predictor, the regularization technique produced lower test errors in five of the seven datasets studied.

Energy-efficient self -organization of wireless acoustic sensor networks for ground target tracking

With the developments in computing and communication technologies, wireless sensor networks have become popular in wide range of application areas such as health, military, environment and habitant monitoring. Moreover, wireless acoustic sensor networks have been widely used for target tracking applications due to their passive nature, reliability and low cost. Traditionally, acoustic sensor arrays built in linear, circular or other regular shapes are used for tracking acoustic sources. The maintaining of relative geometry of the acoustic sensors in the array is vital for accurate target tracking, which greatly reduces the flexibility of the sensor network. To overcome this limitation, we propose using only a single acoustic sensor at each sensor node. This design greatly improves the flexibility of the sensor network and makes it possible to deploy the sensor network in remote or hostile regions through air-drop or other stealth approaches. Acoustic arrays are capable of performing the target localization or generating the bearing estimations on their own. However, with only a single acoustic sensor, the sensor nodes will not be able to generate such measurements. Thus, self-organization of sensor nodes into virtual arrays to perform the target localization is essential. We developed an energy-efficient and distributed self-organization algorithm for target tracking using wireless acoustic sensor networks. The major error sources of the localization process were studied, and an energy-aware node selection criterion was developed to minimize the target localization errors. Using this node selection criterion, the self-organization algorithm selects a near-optimal localization sensor group to minimize the target tracking errors. In addition, a message passing protocol was developed to implement the self-organization algorithm in a distributed manner. In order to achieve extended sensor network lifetime, energy conservation was incorporated into the self-organization algorithm by incorporating a sleep-wakeup management mechanism with a novel cross layer adaptive wakeup probability adjustment scheme. The simulation results confirm that the developed self-organization algorithm provides satisfactory target tracking performance. Moreover, the energy saving analysis confirms the effectiveness of the cross layer power management scheme in achieving extended sensor network lifetime without degrading the target tracking performance.

Energy-Efficient Self-Organization of Wireless Acoustic Sensor Networks for Ground Target Tracking

With the developments in computing and communication technologies, wireless sensor networks have become popular in wide range of application areas such as health, military, environment and habitant monitoring. Moreover, wireless acoustic sensor networks have been widely used for target tracking applications due to their passive nature, reliability and low cost. Traditionally, acoustic sensor arrays built in linear, circular or other regular shapes are used for tracking acoustic sources. The maintaining of relative geometry of the acoustic sensors in the array is vital for accurate target tracking, which greatly reduces the flexibility of the sensor network. To overcome this limitation, we propose using only a single acoustic sensor at each sensor node. This design greatly improves the flexibility of the sensor network and makes it possible to deploy the sensor network in remote or hostile regions through air-drop or other stealth approaches. Acoustic arrays are capable of performing the target localization or generating the bearing estimations on their own. However, with only a single acoustic sensor, the sensor nodes will not be able to generate such measurements. Thus, self-organization of sensor nodes into virtual arrays to perform the target localization is essential. We developed an energy-efficient and distributed self-organization algorithm for target tracking using wireless acoustic sensor networks. The major error sources of the localization process were studied, and an energy-aware node selection criterion was developed to minimize the target localization errors. Using this node selection criterion, the self-organization algorithm selects a near-optimal localization sensor group to minimize the target tracking errors. In addition, a message passing protocol was developed to implement the self-organization algorithm in a distributed manner. In order to achieve extended sensor network lifetime, energy conservation was incorporated into the self-organization algorithm by incorporating a sleep-wakeup management mechanism with a novel cross layer adaptive wakeup probability adjustment scheme. The simulation results confirm that the developed self-organization algorithm provides satisfactory target tracking performance. Moreover, the energy saving analysis confirms the effectiveness of the cross layer power management scheme in achieving extended sensor network lifetime without degrading the target tracking performance.

Labeling large scale social media data using budget-driven One-class SVM classification

The social media classification problems draw more and more attention in the past few years. With the rapid development of Internet and the popularity of computers, there is astronomical amount of information in the social network (social media platforms). The datasets are generally large scale and are often corrupted by noise. The presence of noise in training set has strong impact on the performance of supervised learning (classification) techniques. A budget-driven One-class SVM approach is presented in this thesis that is suitable for large scale social media data classification. Our approach is based on an existing online One-class SVM learning algorithm, referred as STOCS (Self-Tuning One-Class SVM) algorithm. To justify our choice, we first analyze the noise-resilient ability of STOCS using synthetic data. The experiments suggest that STOCS is more robust against label noise than several other existing approaches. Next, to handle big data classification problem for social media data, we introduce several budget driven features, which allow the algorithm to be trained within limited time and under limited memory requirement. Besides, the resulting algorithm can be easily adapted to changes in dynamic data with minimal computational cost. Compared with two state-of-the-art approaches, Lib-Linear and kNN, our approach is shown to be competitive with lower requirements of memory and time.

Spam campaign detection, analysis, and formalization

Les courriels Spams (courriels indésirables ou pourriels) imposent des coûts annuels extrêmement lourds en termes de temps, d’espace de stockage et d’argent aux utilisateurs privés et aux entreprises. Afin de lutter efficacement contre le problème des spams, il ne suffit pas d’arrêter les messages de spam qui sont livrés à la boîte de réception de l’utilisateur. Il est obligatoire, soit d’essayer de trouver et de persécuter les spammeurs qui, généralement, se cachent derrière des réseaux complexes de dispositifs infectés, ou d’analyser le comportement des spammeurs afin de trouver des stratégies de défense appropriées. Cependant, une telle tâche est difficile en raison des techniques de camouflage, ce qui nécessite une analyse manuelle des spams corrélés pour trouver les spammeurs. Pour faciliter une telle analyse, qui doit être effectuée sur de grandes quantités des courriels non classés, nous proposons une méthodologie de regroupement catégorique, nommé CCTree, permettant de diviser un grand volume de spams en des campagnes, et ce, en se basant sur leur similarité structurale. Nous montrons l’efficacité et l’efficience de notre algorithme de clustering proposé par plusieurs expériences. Ensuite, une approche d’auto-apprentissage est proposée pour étiqueter les campagnes de spam en se basant sur le but des spammeur, par exemple, phishing. Les campagnes de spam marquées sont utilisées afin de former un classificateur, qui peut être appliqué dans la classification des nouveaux courriels de spam. En outre, les campagnes marquées, avec un ensemble de quatre autres critères de classement, sont ordonnées selon les priorités des enquêteurs. Finalement, une structure basée sur le semiring est proposée pour la représentation abstraite de CCTree. Le schéma abstrait de CCTree, nommé CCTree terme, est appliqué pour formaliser la parallélisation du CCTree. Grâce à un certain nombre d’analyses mathématiques et de résultats expérimentaux, nous montrons l’efficience et l’efficacité du cadre proposé.

Hybrid mathematical and informational modeling of beam-to-column connections

The analysis of steel and composite frames has traditionally been carried out by idealizing beam-to-column connections as either rigid or pinned. Although some advanced analysis methods have been proposed to account for semi-rigid connections, the performance of these methods strongly depends on the proper modeling of connection behavior. The primary challenge of modeling beam-to-column connections is their inelastic response and continuously varying stiffness, strength, and ductility. In this dissertation, two distinct approaches—mathematical models and informational models—are proposed to account for the complex hysteretic behavior of beam-to-column connections. The performance of the two approaches is examined and is then followed by a discussion of their merits and deficiencies. To capitalize on the merits of both mathematical and informational representations, a new approach, a hybrid modeling framework, is developed and demonstrated through modeling beam-to-column connections. Component-based modeling is a compromise spanning two extremes in the field of mathematical modeling: simplified global models and finite element models. In the component-based modeling of angle connections, the five critical components of excessive deformation are identified. Constitutive relationships of angles, column panel zones, and contact between angles and column flanges, are derived by using only material and geometric properties and theoretical mechanics considerations. Those of slip and bolt hole ovalization are simplified by empirically-suggested mathematical representation and expert opinions. A mathematical model is then assembled as a macro-element by combining rigid bars and springs that represent the constitutive relationship of components. Lastly, the moment-rotation curves of the mathematical models are compared with those of experimental tests. In the case of a top-and-seat angle connection with double web angles, a pinched hysteretic response is predicted quite well by complete mechanical models, which take advantage of only material and geometric properties. On the other hand, to exhibit the highly pinched behavior of a top-and-seat angle connection without web angles, a mathematical model requires components of slip and bolt hole ovalization, which are more amenable to informational modeling. An alternative method is informational modeling, which constitutes a fundamental shift from mathematical equations to data that contain the required information about underlying mechanics. The information is extracted from observed data and stored in neural networks. Two different training data sets, analytically-generated and experimental data, are tested to examine the performance of informational models. Both informational models show acceptable agreement with the moment-rotation curves of the experiments. Adding a degradation parameter improves the informational models when modeling highly pinched hysteretic behavior. However, informational models cannot represent the contribution of individual components and therefore do not provide an insight into the underlying mechanics of components. In this study, a new hybrid modeling framework is proposed. In the hybrid framework, a conventional mathematical model is complemented by the informational methods. The basic premise of the proposed hybrid methodology is that not all features of system response are amenable to mathematical modeling, hence considering informational alternatives. This may be because (i) the underlying theory is not available or not sufficiently developed, or (ii) the existing theory is too complex and therefore not suitable for modeling within building frame analysis. The role of informational methods is to model aspects that the mathematical model leaves out. Autoprogressive algorithm and self-learning simulation extract the missing aspects from a system response. In a hybrid framework, experimental data is an integral part of modeling, rather than being used strictly for validation processes. The potential of the hybrid methodology is illustrated through modeling complex hysteretic behavior of beam-to-column connections. Mechanics-based components of deformation such as angles, flange-plates, and column panel zone, are idealized to a mathematical model by using a complete mechanical approach. Although the mathematical model represents envelope curves in terms of initial stiffness and yielding strength, it is not capable of capturing the pinching effects. Pinching is caused mainly by separation between angles and column flanges as well as slip between angles/flange-plates and beam flanges. These components of deformation are suitable for informational modeling. Finally, the moment-rotation curves of the hybrid models are validated with those of the experimental tests. The comparison shows that the hybrid models are capable of representing the highly pinched hysteretic behavior of beam-to-column connections. In addition, the developed hybrid model is successfully used to predict the behavior of a newly-designed connection.

Aplicaciones de los modelos de predicción bio-inspirados en la administración

Las organizaciones y sus entornos son sistemas complejos. Tales sistemas son difíciles de comprender y predecir. Pese a ello, la predicción es una tarea fundamental para la gestión empresarial y para la toma de decisiones que implica siempre un riesgo. Los métodos clásicos de predicción (entre los cuales están: la regresión lineal, la Autoregresive Moving Average y el exponential smoothing) establecen supuestos como la linealidad, la estabilidad para ser matemática y computacionalmente tratables. Por diferentes medios, sin embargo, se han demostrado las limitaciones de tales métodos. Pues bien, en las últimas décadas nuevos métodos de predicción han surgido con el fin de abarcar la complejidad de los sistemas organizacionales y sus entornos, antes que evitarla. Entre ellos, los más promisorios son los métodos de predicción bio-inspirados (ej. redes neuronales, algoritmos genéticos /evolutivos y sistemas inmunes artificiales). Este artículo pretende establecer un estado situacional de las aplicaciones actuales y potenciales de los métodos bio-inspirados de predicción en la administración.