Search

It has become increasingly popular to obtain machine learning labels through commercial crowdsourcing services. The crowdsourcing workers or annotators are paid for each label they provide, but the task requester usually has only a limited amount of the budget. Since the data instances have different levels of labeling difficulty and the workers have different reliability for the labeling task, it is desirable to wisely allocate the budget among all the instances and workers such that the overall labeling quality is maximized. In this paper, we formulate the budget allocation problem as a Bayesian Markov decision process (MDP), which simultaneously conducts learning and decision making. The optimal allocation policy can be obtained by using the dynamic programming (DP) recurrence. However, DP quickly becomes computationally intractable when the size of the problem increases. To solve this challenge, we propose a computationally efficient approximate policy which is called optimistic knowledge gradient. Our method applies to both pull crowdsourcing marketplaces with homogeneous workers and push marketplaces with heterogeneous workers. It can also incorporate the contextual information of instances when they are available. The experiments on both simulated and real data show that our policy achieves a higher labeling quality than other existing policies at the same budget level. [ABSTRACT FROM AUTHOR]

Hierarchical statistical models are widely employed in information science and data engineering. The models consist of two types of variables: observable variables that represent the given data and latent variables for the unobservable labels. An asymptotic analysis of the models plays an important role in evaluating the learning process; the result of the analysis is applied not only to theoretical but also to practical situations, such as optimal model selection and active learning. There are many studies of generalization errors, which measure the prediction accuracy of the observable variables. However, the accuracy of estimating the latent variables has not yet been elucidated. For a quantitative evaluation of this, the present paper formulates distribution-based functions for the errors in the estimation of the latent variables. The asymptotic behavior is analyzed for both the maximum likelihood and the Bayes methods. [ABSTRACT FROM AUTHOR]

Decision forests are an increasingly popular tool in computer vision problems. Their ad- vantages include high computational efficiency, state-of-the-art accuracy and multi-class support. In this paper, we present a novel method for transfer learning which uses decision forests, and we apply it to recognize gestures and characters. We introduce two mecha- nisms into the decision forest framework in order to transfer knowledge from the source tasks to a given target task. The first one is mixed information gain, which is a data- based regularizer. The second one is label propagation, which infers the manifold structure of the feature space. We show that both of them are important to achieve higher accu- racy. Our experiments demonstrate improvements over traditional decision forests in the ChaLearn Gesture Challenge and MNIST data set. They also compare favorably against other state-of-the-art classiffiers. [ABSTRACT FROM AUTHOR]

Trace-norm regularization is a widely-used and successful approach for collaborative filtering and matrix completion. However, previous learning guarantees require strong assumptions, such as a uniform distribution over the matrix entries. In this paper, we bridge this gap by providing such guarantees, under much milder assumptions which correspond to matrix completion as performed in practice. In fact, we claim that previous difficulties partially stemmed from a mismatch between the standard learning-theoretic modeling of matrix completion, and its practical application. Our results also shed some light on the issue of matrix completion with bounded models, which enforce predictions to lie within a certain range. In particular, we provide experimental and theoretical evidence that such models lead to a modest yet significant improvement. [ABSTRACT FROM AUTHOR]

This paper is concerned with various combinatorial parameters of classes that can be learned from a small set of examples. We show that the recursive teaching dimension, recently introduced by Zilles et al. (2008), is strongly connected to known complexity notions in machine learning, e.g., the self-directed learning complexity and the VC-dimension. To the best of our knowledge these are the first results unveiling such relations between teaching and query learning as well as between teaching and the VC-dimension. It will turn out that for many natural classes the RTD is upper-bounded by the VCD, e.g., classes of VC-dimension 1, intersection-closed classes and finite maximum classes. However, we will also show that there are certain (but rare) classes for which the recursive teaching dimension exceeds the VC-dimension. Moreover, for maximum classes, the combinatorial structure induced by the RTD, called teaching plan, is highly similar to the structure of sample compression schemes. Indeed one can transform any repetition-free teaching plan for a maximum class C into an unlabeled sample compression scheme for C and vice versa, while the latter is produced by (i) the corner-peeling algorithm of Rubinstein and Rubinstein (2012) and (ii) the tail matching algorithm of Kuzmin and Warmuth (2007). [ABSTRACT FROM AUTHOR]

The purpose of this paper is to describe one-shot-learning gesture recognition systems developed on the ChaLearn Gesture Dataset (ChaLearn). We use RGB and depth images and combine appearance (Histograms of Oriented Gradients) and motion descriptors (Histogram of Optical Flow) for parallel temporal segmentation and recognition. The Quadratic-Chi distance family is used to measure differences between histograms to capture cross-bin relationships. We also propose a new algorithm for trimming videos--to remove all the unimportant frames from videos. We present two methods that use a combination of HOG-HOF descriptors together with variants of a Dynamic Time Warping technique. Both methods outperform other published methods and help narrow the gap between human performance and algorithms on this task. The code is publicly available in the MLOSS repository. [ABSTRACT FROM AUTHOR]

This paper proposes an online tree-based Bayesian approach for reinforcement learning. For inference, we employ a generalised context tree model. This defines a distribution on multivariate Gaussian piecewise-linear models, which can be updated in closed form. The tree structure itself is constructed using the cover tree method, which remains efficient in high dimensional spaces. We combine the model with Thompson sampling and approximate dynamic programming to obtain effective exploration policies in unknown environments. The flexibility and computational simplicity of the model render it suitable for many reinforcement learning problems in continuous state spaces. We demonstrate this in an experimental comparison with a Gaussian process model, a linear model and simple least squares policy iteration. [ABSTRACT FROM AUTHOR]

Machine learning algorithms have successfully entered industry through many real-world applications (e.g., search engines and product recommendations). In these applications, the test-time CPU cost must be budgeted and accounted for. In this paper, we examine two main components of the test-time CPU cost, classifier evaluation cost and feature extraction cost, and show how to balance these costs with the classifier accuracy. Since the computation required for feature extraction dominates the test-time cost of a classifier in these settings, we develop two algorithms to efficiently balance the performance with the test-time cost. Our first contribution describes how to construct and optimize a tree of classifiers, through which test inputs traverse along individual paths. Each path extracts different features and is optimized for a specific sub-partition of the input space. Our second contribution is a natural reduction of the tree of classifiers into a cascade. The cascade is particularly useful for class-imbalanced data sets as the majority of instances can be early-exited out of the cascade when the algorithm is sufficiently confident in its prediction. Because both approaches only compute features for inputs that benefit from them the most, we find our trained classifiers lead to high accuracies at a small fraction of the computational cost. [ABSTRACT FROM AUTHOR]

Community detection is the task of detecting hidden communities from observed interactions. Guaranteed community detection has so far been mostly limited to models with non-overlapping communities such as the stochastic block model. In this paper, we remove this restriction, and provide guaranteed community detection for a family of probabilistic network models with overlapping communities, termed as the mixed membership Dirichlet model, first introduced by Airoldi et al. (2008). This model allows for nodes to have fractional memberships in multiple communities and assumes that the community memberships are drawn from a Dirichlet distribution. Moreover, it contains the stochastic block model as a special case. We propose a unified approach to learning these models via a tensor spectral decomposition method. Our estimator is based on low-order moment tensor of the observed network, consisting of 3-star counts. Our learning method is fast and is based on simple linear algebraic operations, e.g., singular value decomposition and tensor power iterations. We provide guaranteed recovery of community memberships and model parameters and present a careful finite sample analysis of our learning method. As an important special case, our results match the best known scaling requirements for the (homogeneous) stochastic block model. [ABSTRACT FROM AUTHOR]

In machine learning, a popular tool to analyze the structure of graphs is the hitting time and the commute distance (resistance distance). For two vertices u and v, the hitting time Huv is the expected time it takes a random walk to travel from u to v. The commute distance is its symmetrized version Cuv = Huv + Hvu. In our paper we study the behavior of hitting times and commute distances when the number n of vertices in the graph tends to infinity. We focus on random geometric graphs (ε-graphs, kNN graphs and Gaussian similarity graphs), but our results also extend to graphs with a given expected degree distribution or Erdõs-Rényi graphs with planted partitions. We prove that in these graph families, the suitably rescaled hitting time Huv converges to 1/ dv and the rescaled commute time to 1/ du + 1/ dv where du and dv denote the degrees of vertices u and v. In these cases, hitting and commute times do not provide information about the structure of the graph, and their use is discouraged in many machine learning applications. [ABSTRACT FROM AUTHOR]

In many machine learning problems such as the dual form of SVM, the objective function to be minimized is convex but not strongly convex. This fact causes difficulties in obtaining the complexity of some commonly used optimization algorithms. In this paper, we proved the global linear convergence on a wide range of algorithms when they are applied to some non-strongly convex problems. In particular, we are the first to prove O(log(1=ϵ)) time complexity of cyclic coordinate descent methods on dual problems of support vector classification and regression. [ABSTRACT FROM AUTHOR]

We consider a framework for structured prediction based on search in the space of complete structured outputs. Given a structured input, an output is produced by running a timebounded search procedure guided by a learned cost function, and then returning the least cost output uncovered during the search. This framework can be instantiated for a wide range of search spaces and search procedures, and easily incorporates arbitrary structuredprediction loss functions. In this paper, we make two main technical contributions. First, we describe a novel approach to automatically defining an effective search space over structured outputs, which is able to leverage the availability of powerful classification learning algorithms. In particular, we define the limited-discrepancy search space and relate the quality of that space to the quality of learned classifiers. We also define a sparse version of the search space to improve the efficiency of our overall approach. Second, we give a generic cost function learning approach that is applicable to a wide range of search procedures. The key idea is to learn a cost function that attempts to mimic the behavior of conducting searches guided by the true loss function. Our experiments on six benchmark domains show that a small amount of search in limited discrepancy search space is often sufficient for significantly improving on state-of-the-art structured-prediction performance. We also demonstrate significant speed improvements for our approach using sparse search spaces with little or no loss in accuracy. [ABSTRACT FROM AUTHOR]

Classical boosting algorithms, such as AdaBoost, build a strong classifier without concern for the computational cost. Some applications, in particular in computer vision, may involve millions of training examples and very large feature spaces. In such contexts, the training time of off-the-shelf boosting algorithms may become prohibitive. Several methods exist to accelerate training, typically either by sampling the features or the examples used to train the weak learners. Even if some of these methods provide a guaranteed speed improvement, they offer no insurance of being more efficient than any other, given the same amount of time. The contributions of this paper are twofold: (1) a strategy to better deal with the increasingly common case where features come from multiple sources (for example, color, shape, texture, etc., in the case of images) and therefore can be partitioned into meaningful subsets; (2) new algorithms which balance at every boosting iteration the number of weak learners and the number of training examples to look at in order to maximize the expected loss reduction. Experiments in image classification and object recognition on four standard computer vision data sets show that the adaptive methods we propose outperform basic sampling and state-of-the-art bandit methods. [ABSTRACT FROM AUTHOR]

Dependencies among neighboring labels in a sequence are important sources of information for sequence labeling and segmentation. However, only first-order dependencies, which are dependencies between adjacent labels or segments, are commonly exploited in practice because of the high computational complexity of typical inference algorithms when longer distance dependencies are taken into account. In this paper, we give efficient inference algorithms to handle high-order dependencies between labels or segments in conditional random fields, under the assumption that the number of distinct label patterns used in the features is small. This leads to efficient learning algorithms for these conditional random fields. We show experimentally that exploiting high-order dependencies can lead to substantial performance improvements for some problems, and we discuss conditions under which high-order features can be effective. [ABSTRACT FROM AUTHOR]

Most existing algorithms for learning Markov network structure either are limited to learning interactions among few variables or are very slow, due to the large space of possible structures. In this paper, we propose three new methods for using decision trees to learn Markov network structures. The advantage of using decision trees is that they are very fast to learn and can represent complex interactions among many variables. The first method, DTSL, learns a decision tree to predict each variable and converts each tree into a set of conjunctive features that define the Markov network structure. The second, DT-BLM, builds on DTSL by using it to initialize a search-based Markov network learning algorithm recently proposed by Davis and Domingos (2010). The third, DT+L1, combines the features learned by DTSL with those learned by an L1-regularized logistic regression method (L1) proposed by Ravikumar et al. (2009). In an extensive empirical evaluation on 20 data sets, DTSL is comparable to L1 and significantly faster and more accurate than two other baselines. DT-BLM is slower than DTSL, but obtains slightly higher accuracy. DT+L1 combines the strengths of DTSL and L1 to perform significantly better than either of them with only a modest increase in training time. [ABSTRACT FROM AUTHOR]

We study a simple linear regression problem for grouped variables; we are interested in methods which jointly perform estimation and variable selection, that is, that automatically set to zero groups of variables in the regression vector. The Group Lasso (GLasso), a well known approach used to tackle this problem which is also a special case of Multiple Kernel Learning (MKL), boils down to solving convex optimization problems. On the other hand, a Bayesian approach commonly known as Sparse Bayesian Learning (SBL), a version of which is the well known Automatic Relevance Determination (ARD), lead to non-convex problems. In this paper we discuss the relation between ARD (and a penalized version which we call PARD) and Glasso, and study their asymptotic properties in terms of the Mean Squared Error in estimating the unknown parameter. The theoretical arguments developed here are independent of the correctness of the prior models and clarify the advantages of PARD over GLasso. [ABSTRACT FROM AUTHOR]

In this paper we establish that the Lovász ϑ function on a graph can be restated as a kernel learning problem. We introduce the notion of SVM - ϑ graphs, on which Lovász ϑ function can be approximated well by a Support vector machine (SVM). We show that Erdös-Rényi random G(n, p) graphs are SVM - ϑ graphs for log4 n/n ≤ p < 1. Even if we embed a large clique of size Θ (√np/1-p) in a G(n, p) graph the resultant graph still remains a SVM - ϑ graph. This immediately suggests an SVM based algorithm for recovering a large planted clique in random graphs. Associated with the ϑ function is the notion of orthogonal labellings. We introduce common orthogonal labellings which extends the idea of orthogonal labellings to multiple graphs. This allows us to propose a Multiple Kernel learning (MKL) based solution which is capable of identifying a large common dense subgraph in multiple graphs. Both in the planted clique case and common subgraph detection problem the proposed solutions beat the state of the art by an order of magnitude. [ABSTRACT FROM AUTHOR]

We present a theoretical analysis for prediction algorithms based on association rules. As part of this analysis, we introduce a problem for which rules are particularly natural, called "sequential event prediction." In sequential event prediction, events in a sequence are revealed one by one, and the goal is to determine which event will next be revealed. The training set is a collection of past sequences of events. An example application is to predict which item will next be placed into a customer's online shopping cart, given his/her past purchases. In the context of this problem, algorithms based on association rules have distinct advantages over classical statistical and machine learning methods: they look at correlations based on subsets of co-occurring past events (items a and b imply item c), they can be applied to the sequential event prediction problem in a natural way, they can potentially handle the "cold start" problem where the training set is small, and they yield interpretable predictions. In this work, we present two algorithms that incorporate association rules. These algorithms can be used both for sequential event prediction and for supervised classification, and they are simple enough that they can possibly be understood by users, customers, patients, managers, etc. We provide generalization guarantees on these algorithms based on algorithmic stability analysis from statistical learning theory. We include a discussion of the strict minimum support threshold often used in association rule mining, and introduce an "adjusted confidence" measure that provides a weaker minimum support condition that has advantages over the strict minimum support. The paper brings together ideas from statistical learning theory, association rule mining and Bayesian analysis. [ABSTRACT FROM AUTHOR]

The Neyman-Pearson (NP) paradigm in binary classification treats type I and type II errors with different priorities. It seeks classifiers that minimize type II error, subject to a type I error constraint under a user specified level α. In this paper, plug-in classifiers are developed under the NP paradigm. Based on the fundamental Neyman-Pearson Lemma, we propose two related plug-in classifiers which amount to thresholding respectively the class conditional density ratio and the regression function. These two classifiers handle different sampling schemes. This work focuses on theoretical properties of the proposed classifiers; in particular, we derive oracle inequalities that can be viewed as finite sample versions of risk bounds. NP classification can be used to address anomaly detection problems, where asymmetry in errors is an intrinsic property. As opposed to a common practice in anomaly detection that consists of thresholding normal class density, our approach does not assume a specific form for anomaly distributions. Such consideration is particularly necessary when the anomaly class density is far from uniformly distributed. [ABSTRACT FROM AUTHOR]

This paper examines the efficacy of sampling-based low-rank approximation techniques when applied to large dense kernel matrices. We analyze two common approximate singular value decomposition techniques, namely the Nyström and Column sampling methods. We present a theoretical comparison between these two methods, provide novel insights regarding their suitability for various tasks and present experimental results that support our theory. Our results illustrate the relative strengths of each method. We next examine the performance of these two techniques on the large-scale task of extracting low-dimensional manifold structure given millions of high-dimensional face images. We address the computational challenges of non-linear dimensionality reduction via Isomap and Laplacian Eigenmaps, using a graph containing about 18 million nodes and 65 million edges. We present extensive experiments on learning low-dimensional embeddings for two large face data sets: CMU-PIE (35 thousand faces) and a web data set (18 million faces). Our comparisons show that the Nyström approximation is superior to the Column sampling method f or this task. Furthermore, approximate Isomap tends to perform better than Laplacian Eigenmaps on both clustering and classification with the labeled CMU-PIE data set. [ABSTRACT FROM AUTHOR]

In this paper, we propose QUANTMINER, a mining quantitative association rules system. This system is based on a genetic algorithm that dynamically discovers "good" intervals in association rules by optimizing both the support and the confidence. The experiments on real and artificial databases have shown the usefulness of QUANTMINER as an interactive, exploratory data mining tool. [ABSTRACT FROM AUTHOR]

Randomized controlled experiments are often described as the most reliable tool available to scientists for discovering causal relationships among quantities of interest. However, it is often unclear how many and which different experiments are needed to identify the full (possibly cyclic) causal structure among some given (possibly causally insufficient) set of variables. Recent results in the causal discovery literature have explored various identifiability criteria that depend on the assumptions one is able to make about the underlying causal process, but these criteria are not directly constructive for selecting the optimal set of experiments. Fortunately, many of the needed constructions already exist in the combinatorics literature, albeit under terminology which is unfamiliar to most of the causal discovery community. In this paper we translate the theoretical results and apply them to the concrete problem of experiment selection. For a variety of settings we give explicit constructions of the optimal set of experiments and adapt some of the general combinatorics results to answer questions relating to the problem of experiment selection. [ABSTRACT FROM AUTHOR]

Recently, researchers have proposed penalized maximum likelihood to identify network topology underlying a dynamical system modeled by multivariate time series. The time series of interest are assumed to be stationary, but this restriction is never taken into consideration by existing estimation methods. Moreover, practical problems of interest may have ultra-high dimensionality and obvious node collinearity. In addition, none of the available algorithms provides a probabilistic measure of the uncertainty for the obtained network topology which is informative in reliable network identification. The main purpose of this paper is to tackle these challenging issues. We propose the S² learning framework, which stands for stationary-sparse network learning. We propose a novel algorithm referred to as the Berhu iterative sparsity pursuit with stationarity (BISPS), where the Berhu regularization can improve the Lasso in detection and estimation. The algorithm is extremely easy to implement, efficient in computation and has a theoretical guarantee to converge to a global optimum. We also incorporate a screening technique into BISPS to tackle ultra-high dimensional problems and enhance computational efficiency. Furthermore, a stationary bootstrap technique is applied to provide connection occurring frequency for reliable topology learning. Experiments show that our method can achieve stationary and sparse causality network learning and is scalable for high-dimensional problems. [ABSTRACT FROM AUTHOR]

Multi-task sparse feature learning aims to improve the generalization performance by exploiting the shared features among tasks. It has been successfully applied to many applications including computer vision and biomedical informatics. Most of the existing multi-task sparse feature learning algorithms are formulated as a convex sparse regularization problem, which is usually suboptimal, due to its looseness for approximating an ℓ0-type regularizer. In this paper, we propose a non-convex formulation for multi-task sparse feature learning based on a novel non-convex regularizer. To solve the non-convex optimization problem, we propose a Multi-Stage Multi-Task Feature Learning (MSMTFL) algorithm; we also provide intuitive interpretations, detailed convergence and reproducibility analysis for the proposed algorithm. Moreover, we present a detailed theoretical analysis showing that MSMTFL achieves a better parameter estimation error bound than the convex formulation. Empirical studies on both synthetic and real-world data sets demonstrate the effectiveness of MSMTFL in comparison with the state of the art multi-task sparse feature learning algorithms. [ABSTRACT FROM AUTHOR]

Imputing missing values in high dimensional time-series is a difficult problem. This paper presents a strategy for training energy-based graphical models for imputation directly, bypassing difficulties probabilistic approaches would face. The training strategy is inspired by recent work on optimization-based learning (Domke, 2012) and allows complex neural models with convolutional and recurrent structures to be trained for imputation tasks. In this work, we use this training strategy to derive learning rules for three substantially different neural architectures. Inference in these models is done by either truncated gradient descent or variational mean-field iterations. In our experiments, we found that the training methods outperform the Contrastive Divergence learning algorithm. Moreover, the training methods can easily handle missing values in the training data itself during learning. We demonstrate the performance of this learning scheme and the three models we introduce on one artificial and two real-world data sets. [ABSTRACT FROM AUTHOR]

We introduce a new class of quadratic support (QS) functions, many of which already play a crucial role in a variety of applications, including machine learning, robust statistical inference, sparsity promotion, and inverse problems such as Kalman smoothing. Well known examples of QS penalties include the l2, Huber, l1 and Vapnik losses. We build on a dual representation for QS functions, using it to characterize conditions necessary to interpret these functions as negative logs of true probability densities. This interpretation establishes the foundation for statistical modeling with both known and new QS loss functions, and enables construction of non-smooth multivariate distributions with specified means and variances from simple scalar building blocks. The main contribution of this paper is a flexible statistical modeling framework for a variety of learning applications, together with a toolbox of efficient numerical methods for estimation. In particular, a broad subclass of QS loss functions known as piecewise linear quadratic (PLQ) penalties has a dual representation that can be exploited to design interior point (IP) methods. IP methods solve nonsmooth optimization problems by working directly with smooth systems of equations characterizing their optimality. We provide several numerical examples, along with a code that can be used to solve general PLQ problems. The efficiency of the IP approach depends on the structure of particular applications. We consider the class of dynamic inverse problems using Kalman smoothing. This class comprises a wide variety of applications, where the aim is to reconstruct the state of a dynamical system with known process and measurement models starting from noisy output samples. In the classical case, Gaussian errors are assumed both in the process and measurement models for such problems. We show that the extended framework allows arbitrary PLQ densities to be used, and that the proposed IP approach solves the generalized Kalman smoothing problem while maintaining the linear complexity in the size of the time series, just as in the Gaussian case. This extends the computational efficiency of the Mayne-Fraser and Rauch-Tung-Striebel algorithms to a much broader nonsmooth setting, and includes many recently proposed robust and sparse smoothers as special cases. [ABSTRACT FROM AUTHOR]

Despite the simplicity of the Naive Bayes classifier, it has continued to perform well against more sophisticated newcomers and has remained, therefore, of great interest to the machine learning community. Of numerous approaches to refining the naive Bayes classifier, attribute weighting has received less attention than it warrants. Most approaches, perhaps influenced by attribute weighting in other machine learning algorithms, use weighting to place more emphasis on highly predictive attributes than those that are less predictive. In this paper, we argue that for naive Bayes attribute weighting should instead be used to alleviate the conditional independence assumption. Based on this premise, we propose a weighted naive Bayes algorithm, called WANBIA, that selects weights to minimize either the negative conditional log likelihood or the mean squared error objective functions. We perform extensive evaluations and find thatWANBIA is a competitive alternative to state of the art classifiers like Random Forest, Logistic Regression and A1DE. [ABSTRACT FROM AUTHOR]

This paper studies the asymptotic behaviors of the pairwise angles among n randomly and uniformly distributed unit vectors in ℝp as the number of points n → ∞, while the dimension p is either fixed or growing with n. For both settings, we derive the limiting empirical distribution of the random angles and the limiting distributions of the extreme angles. The results reveal interesting differences in the two settings and provide a precise characterization of the folklore that "all high-dimensional random vectors are almost always nearly orthogonal to each other". Applications to statistics and machine learning and connections with some open problems in physics and mathematics are also discussed. [ABSTRACT FROM AUTHOR]

Constraint-based learning of Bayesian networks (BN) from limited data can lead to multiple testing problems when recovering dense areas of the skeleton and to conflicting results in the orientation of edges. In this paper, we present a new constraint-based algorithm, light mutual min (LMM) for improved accuracy of BN learning from small sample data. LMM improves the assessment of candidate edges by using a ranking criterion that considers conditional independence on neighboring variables at both sides of an edge simultaneously. The algorithm also employs an adaptive relaxation of constraints that, selectively, allows some nodes not to condition on some neighbors. This relaxation aims at reducing the incorrect rejection of true edges connecting high degree nodes due to multiple testing. LMM additionally incorporates a new criterion for ranking v-structures that is used to recover the completed partially directed acyclic graph (CPDAG) and to resolve conflicting v-structures, a common problem in small sample constraint-based learning. Using simulated data, each of these components of LMM is shown to significantly improve network inference compared to commonly applied methods when learning from limited data, including more accurate recovery of skeletons and CPDAGs compared to the PC, MaxMin, and MaxMin hill climbing algorithms. A proof of asymptotic correctness is also provided for LMM for recovering the correct skeleton and CPDAG. [ABSTRACT FROM AUTHOR]

There are two main approaches to binary classification problems: the loss function approach and the uncertainty set approach. The loss function approach is widely used in real-world data analysis. Statistical decision theory has been used to elucidate its properties such as statistical consistency. Conditional probabilities can also be estimated by using the minimum solution of the loss function. In the uncertainty set approach, an uncertainty set is defined for each binary label from training samples. The best separating hyperplane between the two uncertainty sets is used as the decision function. Although the uncertainty set approach provides an intuitive understanding of learning algorithms, its statistical properties have not been sufficiently studied. In this paper, we show that the uncertainty set is deeply connected with the convex conjugate of a loss function. On the basis of the conjugate relation, we propose a way of revising the uncertainty set approach so that it will have good statistical properties such as statistical consistency. We also introduce statistical models corresponding to uncertainty sets in order to estimate conditional probabilities. Finally, we present numerical experiments, verifying that the learning with revised uncertainty sets improves the prediction accuracy. [ABSTRACT FROM AUTHOR]

Principal curves and manifolds provide a framework to formulate manifold learning within a statistical context. Principal curves define the notion of a curve passing through the middle of a distribution. While the intuition is clear, the formal definition leads to some technical and practical difficulties. In particular, principal curves are saddle points of the mean-squared projection distance, which poses severe challenges for estimation and model selection. This paper demonstrates that the difficulties in model selection associated with the saddle point property of principal curves are intrinsically tied to the minimization of the mean-squared projection distance. We introduce a new objective function, facilitated through a modification of the principal curve estimation approach, for which all critical points are principal curves and minima. Thus, the new formulation removes the fundamental issue for model selection in principal curve estimation. A gradient-descent-based estimator demonstrates the effectiveness of the new formulation for controlling model complexity on numerical experiments with synthetic and real data. [ABSTRACT FROM AUTHOR]

A statistical model or a learning machine is called regular if the map taking a parameter to a probability distribution is one-to-one and if its Fisher information matrix is always positive definite. If otherwise, it is called singular. In regular statistical models, the Bayes free energy, which is defined by the minus logarithm of Bayes marginal likelihood, can be asymptotically approximated by the Schwarz Bayes information criterion (BIC), whereas in singular models such approximation does not hold. Recently, it was proved that the Bayes free energy of a singular model is asymptotically given by a generalized formula using a birational invariant, the real log canonical threshold (RLCT), instead of half the number of parameters in BIC. Theoretical values of RLCTs in several statistical models are now being discovered based on algebraic geometrical methodology. However, it has been difficult to estimate the Bayes free energy using only training samples, because an RLCT depends on an unknown true distribution. In the present paper, we define a widely applicable Bayesian information criterion (WBIC) by the average log likelihood function over the posterior distribution with the inverse temperature 1/logn, where n is the number of training samples. We mathematically prove that WBIC has the same asymptotic expansion as the Bayes free energy, even if a statistical model is singular for or unrealizable by a statistical model. Since WBIC can be numerically calculated without any information about a true distribution, it is a generalized version of BIC onto singular statistical models. [ABSTRACT FROM AUTHOR]

Stochastic Gradient Descent (SGD) has become popular for solving large scale supervised machine learning optimization problems such as SVM, due to their strong theoretical guarantees. While the closely related Dual Coordinate Ascent (DCA) method has been implemented in various software packages, it has so far lacked good convergence analysis. This paper presents a new analysis of Stochastic Dual Coordinate Ascent (SDCA) showing that this class of methods enjoy strong theoretical guarantees that are comparable or better than SGD. This analysis justifies the effectiveness of SDCA for practical applications. [ABSTRACT FROM AUTHOR]

Lévy processes refer to a class of stochastic processes, for example, Poisson processes and Brownian motions, and play an important role in stochastic processes and machine learning. Therefore, it is essential to study risk bounds of the learning process for time-dependent samples drawn from a Levy process (or briefly called learning process for Lévy process). It is noteworthy that samples in this learning process are not independently and identically distributed (i.i.d.). Therefore, results in traditional statistical learning theory are not applicable (or at least cannot be applied directly), because they are obtained under the sample-i.i.d. assumption. In this paper, we study risk bounds of the learning process for time-dependent samples drawn from a Lévy process, and then analyze the asymptotical behavior of the learning process. In particular, we first develop the deviation inequalities and the symmetrization inequality for the learning process. By using the resultant inequalities, we then obtain the risk bounds based on the covering number. Finally, based on the resulting risk bounds, we study the asymptotic convergence and the rate of convergence of the learning process for Lévy process. Meanwhile, we also give a comparison to the related results under the sample-i.i.d. assumption. [ABSTRACT FROM AUTHOR]

Boosting combines weak classifiers to form highly accurate predictors. Although the case of binary classification is well understood, in the multiclass setting, the "correct" requirements on the weak classifier, or the notion of the most efficient boosting algorithms are missing. In this paper, we create a broad and general framework, within which we make precise and identify the optimal requirements on the weak-classifier, as well as design the most effective, in a certain sense, boosting algorithms that assume such requirements. [ABSTRACT FROM AUTHOR]

Algorithms for Markov boundary discovery from data constitute an important recent development in machine learning, primarily because they offer a principled solution to the variable/feature selection problem and give insight on local causal structure. Over the last decade many sound algorithms have been proposed to identify a single Markov boundary of the response variable. Even though faithful distributions and, more broadly, distributions that satisfy the intersection property always have a single Markov boundary, other distributions/data sets may have multiple Markov boundaries of the response variable. The latter distributions/data sets are common in practical data-analytic applications, and there are several reasons why it is important to induce multiple Markov boundaries from such data. However, there are currently no sound and efficient algorithms that can accomplish this task. This paper describes a family of algorithms TIE* that can discover all Markov boundaries in a distribution. The broad applicability as well as efficiency of the new algorithmic family is demonstrated in an extensive benchmarking study that involved comparison with 26 state-of-the-art algorithms/variants in 15 data sets from a diversity of application domains. [ABSTRACT FROM AUTHOR]

There is much interest in the Hierarchical Dirichlet Process Hidden Markov Model (HDP-HMM) as a natural Bayesian nonparametric extension of the ubiquitous Hidden Markov Model for learning from sequential and time-series data. However, in many settings the HDP-HMM's strict Markovian constraints are undesirable, particularly if we wish to learn or encode non-geometric state durations. We can extend the HDP-HMM to capture such structure by drawing upon explicit-duration semi-Markov modeling, which has been developed mainly in the parametric non-Bayesian setting, to allow construction of highly interpretable models that admit natural prior information on state durations. In this paper we introduce the explicit-duration Hierarchical Dirichlet Process Hidden semi-Markov Model (HDP-HSMM) and develop sampling algorithms for efficient posterior inference. The methods we introduce also provide new methods for sampling inference in the finite Bayesian HSMM. Our modular Gibbs sampling methods can be embedded in samplers for larger hierarchical Bayesian models, adding semi-Markov chain modeling as another tool in the Bayesian inference toolbox. We demonstrate the utility of the HDP-HSMM and our inference methods on both synthetic and real experiments. [ABSTRACT FROM AUTHOR]

This paper introduces the rllib as an original C++ template-based library oriented toward value function estimation. Generic programming is promoted here as a way of having a good fit between the mathematics of reinforcement learning and their implementation in a library. The main concepts of rllib are presented, as well as a short example. [ABSTRACT FROM AUTHOR]

We consider the problem of parsing human poses and recognizing their actions in static images with part-based models. Most previous work in part-based models only considers rigid parts (e.g., torso, head, half limbs) guided by human anatomy. We argue that this representation of parts is not necessarily appropriate. In this paper, we introduce hierarchical poselets--a new representation for modeling the pose configuration of human bodies. Hierarchical poselets can be rigid parts, but they can also be parts that cover large portions of human bodies (e.g., torso + left arm). In the extreme case, they can be the whole bodies. The hierarchical poselets are organized in a hierarchical way via a structured model. Human parsing can be achieved by inferring the optimal labeling of this hierarchical model. The pose information captured by this hierarchical model can also be used as a intermediate representation for other high-level tasks. We demonstrate it in action recognition from static images. [ABSTRACT FROM AUTHOR]

In this paper, we report a performance bound for the widely used least-squares policy iteration (LSPI) algorithm. We first consider the problem of policy evaluation in reinforcement learning, that is, learning the value function of a fixed policy, using the least-squares temporal-difference (LSTD) learning method, and report finite-sample analysis for this algorithm. To do so, we first derive a bound on the performance of the LSTD solution evaluated at the states generated by the Markov chain and used by the algorithm to learn an estimate of the value function. This result is general in the sense that no assumption is made on the existence of a stationary distribution for the Markov chain. We then derive generalization bounds in the case when the Markov chain possesses a stationary distribution and is β-mixing. Finally, we analyze how the error at each policy evaluation step is propagated through the iterations of a policy iteration method, and derive a performance bound for the LSPI algorithm. [ABSTRACT FROM AUTHOR]

In hierarchical classification, class labels are structured, that is each label value corresponds to one non-root node in a tree, where the inter-class relationship for classification is specified by directed paths of the tree. In such a situation, the focus has been on how to leverage the interclass relationship to enhance the performance of flat classification, which ignores such dependency. This is critical when the number of classes becomes large relative to the sample size. This paper considers single-path or partial-path hierarchical classification, where only one path is permitted from the root to a leaf node. A large margin method is introduced based on a new concept of generalized margins with respect to hierarchy. For implementation, we consider support vector machines and ψ-learning. Numerical and theoretical analyses suggest that the proposed method achieves the desired objective and compares favorably against strong competitors in the literature, including its flat counterparts. Finally, an application to gene function prediction is discussed. [ABSTRACT FROM AUTHOR]

Co-training (or more generally, co-regularization) has been a popular algorithm for semi-supervised learning in data with two feature representations (or views), but the fundamental assumptions underlying this type of models are still unclear. In this paper we propose a Bayesian undirected graphical model for co-training, or more generally for semi-supervised multi-view learning. This makes explicit the previously unstated assumptions of a large class of co-training type algorithms, and also clarifies the circumstances under which these assumptions fail. Building upon new insights from this model, we propose an improved method for co-training, which is a novel co-training kernel for Gaussian process classifiers. The resulting approach is convex and avoids local-maxima problems, and it can also automatically estimate how much each view should be trusted to accommodate noisy or unreliable views. The Bayesian co-training approach can also elegantly handle data samples with missing views, that is, some of the views are not available for some data points at learning time. This is further extended to an active sensing framework, in which the missing (sample, view) pairs are actively acquired to improve learning performance. The strength of active sensing model is that one actively sensed (sample, view) pair would improve the joint multi-view classification on all the samples. Experiments on toy data and several real world data sets illustrate the benefits of this approach. [ABSTRACT FROM AUTHOR]

When training deep networks it is common knowledge that an efficient and well generalizing representation of the problem is formed. In this paper we aim to elucidate what makes the emerging representation successful. We analyze the layer-wise evolution of the representation in a deep network by building a sequence of deeper and deeper kernels that subsume the mapping performed by more and more layers of the deep network and measuring how these increasingly complex kernels fit the learning problem. We observe that deep networks create increasingly better representations of the learning problem and that the structure of the deep network controls how fast the representation of the task is formed layer after layer. [ABSTRACT FROM AUTHOR]

The second eigenvalue of the Laplacian matrix and its associated eigenvector are fundamental features of an undirected graph, and as such they have found widespread use in scientific computing, machine learning, and data analysis. In many applications, however, graphs that arise have several local regions of interest, and the second eigenvector will typically fail to provide information fine-tuned to each local region. In this paper, we introduce a locally-biased analogue of the second eigenvector, and we demonstrate its usefulness at highlighting local properties of data graphs in a semi-supervised manner. To do so, we first view the second eigenvector as the solution to a constrained optimization problem, and we incorporate the local information as an additional constraint; we then characterize the optimal solution to this new problem and show that it can be interpreted as a generalization of a Personalized PageRank vector; and finally, as a consequence, we show that the solution can be computed in nearly-linear time. In addition, we show that this locally-biased vector can be used to compute an approximation to the best partition near an input seed set in a manner analogous to the way in which the second eigenvector of the Laplacian can be used to obtain an approximation to the best partition in the entire input graph. Such a primitive is useful for identifying and refining clusters locally, as it allows us to focus on a local region of interest in a semi-supervised manner. Finally, we provide a detailed empirical evaluation of our method by showing how it can applied to finding locally-biased sparse cuts around an input vertex seed set in social and information networks. [ABSTRACT FROM AUTHOR]

A supervised topic model can use side information such as ratings or labels associated with documents or images to discover more predictive low dimensional topical representations of the data. However, existing supervised topic models predominantly employ likelihood-driven objective functions for learning and inference, leaving the popular and potentially powerful max-margin principle unexploited for seeking predictive representations of data and more discriminative topic bases for the corpus. In this paper, we propose the maximum entropy discrimination latent Dirichlet allocation (MedLDA) model, which integrates the mechanism behind the max-margin prediction models (e.g., SVMs) with the mechanism behind the hierarchical Bayesian topic models (e.g., LDA) under a unified constrained optimization framework, and yields latent topical representations that are more discriminative and more suitable for prediction tasks such as document classification or regression. The principle underlying the MedLDA formalism is quite general and can be applied for jointly max-margin and maximum likelihood learning of directed or undirected topic models when supervising side information is available. Efficient variational methods for posterior inference and parameter estimation are derived and extensive empirical studies on several real data sets are also provided. Our experimental results demonstrate qualitatively and quantitatively that MedLDA could: 1) discover sparse and highly discriminative topical representations; 2) achieve state of the art prediction performance; and 3) be more efficient than existing supervised topic models, especially for classification. [ABSTRACT FROM AUTHOR]

The investigation of directed acyclic graphs (DAGs) encoding the same Markov property, that is the same conditional independence relations of multivariate observational distributions, has a long tradition; many algorithms exist for model selection and structure learning in Markov equivalence classes. In this paper, we extend the notion of Markov equivalence of DAGs to the case of interventional distributions arising from multiple intervention experiments. We show that under reasonable assumptions on the intervention experiments, interventional Markov equivalence defines a finer partitioning of DAGs than observational Markov equivalence and hence improves the identifiability of causal models. We give a graph theoretic criterion for two DAGs being Markov equivalent under interventions and show that each interventional Markov equivalence class can, analogously to the observational case, be uniquely represented by a chain graph called interventional essential graph (also known as CPDAG in the observational case). These are key insights for deriving a generalization of the Greedy Equivalence Search algorithm aimed at structure learning from interventional data. This new algorithm is evaluated in a simulation study. [ABSTRACT FROM AUTHOR]

In this paper we propose a novel framework for the construction of sparsity-inducing priors. In particular, we define such priors as a mixture of exponential power distributions with a generalized inverse Gaussian density (EP-GIG). EP-GIG is a variant of generalized hyperbolic distributions, and the special cases include Gaussian scale mixtures and Laplace scale mixtures. Furthermore, Laplace scale mixtures can subserve a Bayesian framework for sparse learning with nonconvex penalization. The densities of EP-GIG can be explicitly expressed. Moreover, the corresponding posterior distribution also follows a generalized inverse Gaussian distribution. We exploit these properties to develop EM algorithms for sparse empirical Bayesian learning. We also show that these algorithms bear an interesting resemblance to iteratively reweighted l2 or l1 methods. Finally, we present two extensions for grouped variable selection and logistic regression. [ABSTRACT FROM AUTHOR]

We consider a class of sparse learning problems in high dimensional feature space regularized by a structured sparsity-inducing norm that incorporates prior knowledge of the group structure of the features. Such problems often pose a considerable challenge to optimization algorithms due to the non-smoothness and non-separability of the regularization term. In this paper, we focus on two commonly adopted sparsity-inducing regularization terms, the overlapping Group Lasso penalty l1/l2-norm and the l1/l∞-norm. We propose a unified framework based on the augmented Lagrangian method, under which problems with both types of regularization and their variants can be efficiently solved. As one of the core building-blocks of this framework, we develop new algorithms using a partial-linearization/splitting technique and prove that the accelerated versions of these algorithms require O(1/√ε) iterations to obtain an ε-optimal solution. We compare the performance of these algorithms against that of the alternating direction augmented Lagrangian and FISTA methods on a collection of data sets and apply them to two real-world problems to compare the relative merits of the two norms. [ABSTRACT FROM AUTHOR]

Recent developments have demonstrated the capacity of restricted Boltzmann machines (RBM) to be powerful generative models, able to extract useful features from input data or construct deep artificial neural networks. In such settings, the RBMonly yields a preprocessing or an initialization for some other model, instead of acting as a complete supervised model in its own right. In this paper, we argue that RBMs can provide a self-contained framework for developing competitive classifiers. We study the Classification RBM (ClassRBM), a variant on the RBM adapted to the classification setting. We study different strategies for training the ClassRBM and show that competitive classification performances can be reached when appropriately combining discriminative and generative training objectives. Since training according to the generative objective requires the computation of a generally intractable gradient, we also compare different approaches to estimating this gradient and address the issue of obtaining such a gradient for problems with very high dimensional inputs. Finally, we describe how to adapt the ClassRBM to two special cases of classification problems, namely semi-supervised and multitask learning. [ABSTRACT FROM AUTHOR]

Metric and kernel learning arise in several machine learning applications. However, most existing metric learning algorithms are limited to learning metrics over low-dimensional data, while existing kernel learning algorithms are often limited to the transductive setting and do not generalize to new data points. In this paper, we study the connections between metric learning and kernel learning that arise when studying metric learning as a linear transformation learning problem. In particular, we propose a general optimization framework for learning metrics via linear transformations, and analyze in detail a special case of our framework-that of minimizing the LogDet divergence subject to linear constraints. We then propose a general regularized framework for learning a kernel matrix, and show it to be equivalent to our metric learning framework. Our theoretical connections between metric and kernel learning have two main consequences: 1) the learned kernel matrix parameterizes a linear transformation kernel function and can be applied inductively to new data points, 2) our result yields a constructive method for kernelizing most existing Mahalanobis metric learning formulations. We demonstrate our learning approach by applying it to large-scale real world problems in computer vision, text mining and semi-supervised kernel dimensionality reduction. [ABSTRACT FROM AUTHOR]