A sequential dual method for structural SVMs

In many real world prediction problems the output is a structured object like a sequence or a tree or a graph. Such problems range from natural language processing to compu- tational biology or computer vision and have been tackled using algorithms, referred to as structured output learning algorithms. We consider the problem of structured classifi- cation. In the last few years, large margin classifiers like sup-port vector machines (SVMs) have shown much promise for structured output learning. The related optimization prob -lem is a convex quadratic program (QP) with a large num-ber of constraints, which makes the problem intractable for large data sets. This paper proposes a fast sequential dual method (SDM) for structural SVMs. The method makes re-peated passes over the training set and optimizes the dual variables associated with one example at a time. The use of additional heuristics makes the proposed method more efficient. We present an extensive empirical evaluation of the proposed method on several sequence learning problems.Our experiments on large data sets demonstrate that the proposed method is an order of magnitude faster than state of the art methods like cutting-plane method and stochastic gradient descent method (SGD). Further, SDM reaches steady state generalization performance faster than the SGD method. The proposed SDM is thus a useful alternative for large scale structured output learning.

A survey of cloud storage facilities

There are many applications such as software for processing customer records in telecom, patient records in hospitals, email processing software accessing a single email in a mailbox etc. which require to access a single record in a database consisting of millions of records. A basic feature of these applications is that they need to access data sets which are very large but simple. Cloud computing provides computing requirements for these kinds of new generation of applications involving very large data sets which cannot possibly be handled efficiently using traditional computing infrastructure. In this paper, we describe storage services provided by three well-known cloud service providers and give a comparison of their features with a view to characterize storage requirements of very large data sets as examples and we hope that it would act as a catalyst for the design of storage services for very large data set requirements in future. We also give a brief overview of other kinds of storage that have come up in the recent past for cloud computing.

Infinite push: a new support vector ranking algorithm that directly optimizes accuracy at the absolute top of the list

Ranking problems have become increasingly important in machine learning and data mining in recent years, with applications ranging from information retrieval and recommender systems to computational biology and drug discovery. In this paper, we describe a new ranking algorithm that directly maximizes the number of relevant objects retrieved at the absolute top of the list. The algorithm is a support vector style algorithm, but due to the different objective, it no longer leads to a quadratic programming problem. Instead, the dual optimization problem involves l1, ∞ constraints; we solve this dual problem using the recent l1, ∞ projection method of Quattoni et al (2009). Our algorithm can be viewed as an l∞-norm extreme of the lp-norm based algorithm of Rudin (2009) (albeit in a support vector setting rather than a boosting setting); thus we refer to the algorithm as the ‘Infinite Push’. Experiments on real-world data sets confirm the algorithm’s focus on accuracy at the absolute top of the list.

Pattern clustering using cooperative game theory

In this paper, we approach the classical problem of clustering using solution concepts from cooperative game theory such as Nucleolus and Shapley value. We formulate the problem of clustering as a characteristic form game and develop a novel algorithm DRAC (Density-Restricted Agglomerative Clustering) for clustering. With extensive experimentation on standard data sets, we compare the performance of DRAC with that of well known algorithms. We show an interesting result that four prominent solution concepts, Nucleolus, Shapley value, Gately point and \tau-value coincide for the defined characteristic form game. This vindicates the choice of the characteristic function of the clustering game and also provides strong intuitive foundation for our approach.

Reconstructing the properties of dark energy using standard sirens

Future space-based gravity wave (GW) experiments such as the Big Bang Observatory (BBO), with their excellent projected, one sigma angular resolution, will measure the luminosity distance to a large number of GW sources to high precision, and the redshift of the single galaxies in the narrow solid angles towards the sources will provide the redshifts of the gravity wave sources. One sigma BBO beams contain the actual source in only 68% of the cases; the beams that do not contain the source may contain a spurious single galaxy, leading to misidentification. To increase the probability of the source falling within the beam, larger beams have to be considered, decreasing the chances of finding single galaxies in the beams. Saini et al. T.D. Saini, S.K. Sethi, and V. Sahni, Phys. Rev. D 81, 103009 (2010)] argued, largely analytically, that identifying even a small number of GW source galaxies furnishes a rough distance-redshift relation, which could be used to further resolve sources that have multiple objects in the angular beam. In this work we further develop this idea by introducing a self-calibrating iterative scheme which works in conjunction with Monte Carlo simulations to determine the luminosity distance to GW sources with progressively greater accuracy. This iterative scheme allows one to determine the equation of state of dark energy to within an accuracy of a few percent for a gravity wave experiment possessing a beam width an order of magnitude larger than BBO (and therefore having a far poorer angular resolution). This is achieved with no prior information about the nature of dark energy from other data sets such as type Ia supernovae, baryon acoustic oscillations, cosmic microwave background, etc. DOI:10.1103/PhysRevD.87.083001

Efficient method of moving shadow detection and vehicle classification

Moving shadow detection and removal from the extracted foreground regions of video frames, aim to limit the risk of misconsideration of moving shadows as a part of moving objects. This operation thus enhances the rate of accuracy in detection and classification of moving objects. With a similar reasoning, the present paper proposes an efficient method for the discrimination of moving object and moving shadow regions in a video sequence, with no human intervention. Also, it requires less computational burden and works effectively under dynamic traffic road conditions on highways (with and without marking lines), street ways (with and without marking lines). Further, we have used scale-invariant feature transform-based features for the classification of moving vehicles (with and without shadow regions), which enhances the effectiveness of the proposed method. The potentiality of the method is tested with various data sets collected from different road traffic scenarios, and its superiority is compared with the existing methods. (C) 2013 Elsevier GmbH. All rights reserved.

A Hybrid Parallel Algorithm for Computing and Tracking Level Set Topology

The contour tree is a topological abstraction of a scalar field that captures evolution in level set connectivity. It is an effective representation for visual exploration and analysis of scientific data. We describe a work-efficient, output sensitive, and scalable parallel algorithm for computing the contour tree of a scalar field defined on a domain that is represented using either an unstructured mesh or a structured grid. A hybrid implementation of the algorithm using the GPU and multi-core CPU can compute the contour tree of an input containing 16 million vertices in less than ten seconds with a speedup factor of upto 13. Experiments based on an implementation in a multi-core CPU environment show near-linear speedup for large data sets.

DSA image registration using non-uniform MRF model and pivotal control points

In order to reduce the motion artifacts in DSA, non-rigid image registration is commonly used before subtracting the mask from the contrast image. Since DSA registration requires a set of spatially non-uniform control points, a conventional MRF model is not very efficient. In this paper, we introduce the concept of pivotal and non-pivotal control points to address this, and propose a non-uniform MRF for DSA registration. We use quad-trees in a novel way to generate the non-uniform grid of control points. Our MRF formulation produces a smooth displacement field and therefore results in better artifact reduction than that of registering the control points independently. We achieve improved computational performance using pivotal control points without compromising on the artifact reduction. We have tested our approach using several clinical data sets, and have presented the results of quantitative analysis, clinical assessment and performance improvement on a GPU. (C) 2013 Elsevier Ltd. All rights reserved.

Game theoretic, decentralized, local information based algorithm for community detection in social graphs

Motivated by the observation that communities in real world social networks form due to actions of rational individuals in networks, we propose a novel game theory inspired algorithm to determine communities in networks. The algorithm is decentralized and only uses local information at each node. We show the efficacy of the proposed algorithm through extensive experimentation on several real world social network data sets.

Differentially private stochastic gradient descent algorithm for multiparty classification

We consider the problem of developing privacy-preserving machine learning algorithms in a dis-tributed multiparty setting. Here different parties own different parts of a data set, and the goal is to learn a classifier from the entire data set with-out any party revealing any information about the individual data points it owns. Pathak et al [7]recently proposed a solution to this problem in which each party learns a local classifier from its own data, and a third party then aggregates these classifiers in a privacy-preserving manner using a cryptographic scheme. The generaliza-tion performance of their algorithm is sensitive to the number of parties and the relative frac-tions of data owned by the different parties. In this paper, we describe a new differentially pri-vate algorithm for the multiparty setting that uses a stochastic gradient descent based procedure to directly optimize the overall multiparty ob-jective rather than combining classifiers learned from optimizing local objectives. The algorithm achieves a slightly weaker form of differential privacy than that of [7], but provides improved generalization guarantees that do not depend on the number of parties or the relative sizes of the individual data sets. Experimental results corrob-orate our theoretical findings.

Probabilistic least squares approach to ordinal regression

This paper proposes a novel approach to solve the ordinal regression problem using Gaussian processes. The proposed approach, probabilistic least squares ordinal regression (PLSOR), obtains the probability distribution over ordinal labels using a particular likelihood function. It performs model selection (hyperparameter optimization) using the leave-one-out cross-validation (LOO-CV) technique. PLSOR has conceptual simplicity and ease of implementation of least squares approach. Unlike the existing Gaussian process ordinal regression (GPOR) approaches, PLSOR does not use any approximation techniques for inference. We compare the proposed approach with the state-of-the-art GPOR approaches on some synthetic and benchmark data sets. Experimental results show the competitiveness of the proposed approach.

Validation based sparse Gaussian processes for ordinal regression

This paper proposes a sparse modeling approach to solve ordinal regression problems using Gaussian processes (GP). Designing a sparse GP model is important from training time and inference time viewpoints. We first propose a variant of the Gaussian process ordinal regression (GPOR) approach, leave-one-out GPOR (LOO-GPOR). It performs model selection using the leave-one-out cross-validation (LOO-CV) technique. We then provide an approach to design a sparse model for GPOR. The sparse GPOR model reduces computational time and storage requirements. Further, it provides faster inference. We compare the proposed approaches with the state-of-the-art GPOR approach on some benchmark data sets. Experimental results show that the proposed approaches are competitive.

Simultaneous feature selection and clustering using particle swarm optimization

Data clustering groups data so that data which are similar to each other are in the same group and data which are dissimilar to each other are in different groups. Since generally clustering is a subjective activity, it is possible to get different clusterings of the same data depending on the need. This paper attempts to find the best clustering of the data by first carrying out feature selection and using only the selected features, for clustering. A PSO (Particle Swarm Optimization)has been used for clustering but feature selection has also been carried out simultaneously. The performance of the above proposed algorithm is evaluated on some benchmark data sets. The experimental results shows the proposed methodology outperforms the previous approaches such as basic PSO and Kmeans for the clustering problem.

A MatLAB Based Virtual Phantom for 2D Electrical Impedance Tomography (MatVP2DEIT): Studying the Medical Electrical Impedance Tomography Reconstruction in Computer

Electrical Impedance Tomography (EIT) is a computerized medical imaging technique which reconstructs the electrical impedance images of a domain under test from the boundary voltage-current data measured by an EIT electronic instrumentation using an image reconstruction algorithm. Being a computed tomography technique, EIT injects a constant current to the patient's body through the surface electrodes surrounding the domain to be imaged (Omega) and tries to calculate the spatial distribution of electrical conductivity or resistivity of the closed conducting domain using the potentials developed at the domain boundary (partial derivative Omega). Practical phantoms are essentially required to study, test and calibrate a medical EIT system for certifying the system before applying it on patients for diagnostic imaging. Therefore, the EIT phantoms are essentially required to generate boundary data for studying and assessing the instrumentation and inverse solvers a in EIT. For proper assessment of an inverse solver of a 2D EIT system, a perfect 2D practical phantom is required. As the practical phantoms are the assemblies of the objects with 3D geometries, the developing of a practical 2D-phantom is a great challenge and therefore, the boundary data generated from the practical phantoms with 3D geometry are found inappropriate for assessing a 2D inverse solver. Furthermore, the boundary data errors contributed by the instrumentation are also difficult to separate from the errors developed by the 3D phantoms. Hence, the errorless boundary data are found essential to assess the inverse solver in 2D EIT. In this direction, a MatLAB-based Virtual Phantom for 2D EIT (MatVP2DEIT) is developed to generate accurate boundary data for assessing the 2D-EIT inverse solvers and the image reconstruction accuracy. MatVP2DEIT is a MatLAB-based computer program which simulates a phantom in computer and generates the boundary potential data as the outputs by using the combinations of different phantom parameters as the inputs to the program. Phantom diameter, inhomogeneity geometry (shape, size and position), number of inhomogeneities, applied current magnitude, background resistivity, inhomogeneity resistivity all are set as the phantom variables which are provided as the input parameters to the MatVP2DEIT for simulating different phantom configurations. A constant current injection is simulated at the phantom boundary with different current injection protocols and boundary potential data are calculated. Boundary data sets are generated with different phantom configurations obtained with the different combinations of the phantom variables and the resistivity images are reconstructed using EIDORS. Boundary data of the virtual phantoms, containing inhomogeneities with complex geometries, are also generated for different current injection patterns using MatVP2DEIT and the resistivity imaging is studied. The effect of regularization method on the image reconstruction is also studied with the data generated by MatVP2DEIT. Resistivity images are evaluated by studying the resistivity parameters and contrast parameters estimated from the elemental resistivity profiles of the reconstructed phantom domain. Results show that the MatVP2DEIT generates accurate boundary data for different types of single or multiple objects which are efficient and accurate enough to reconstruct the resistivity images in EIDORS. The spatial resolution studies show that, the resistivity imaging conducted with the boundary data generated by MatVP2DEIT with 2048 elements, can reconstruct two circular inhomogeneities placed with a minimum distance (boundary to boundary) of 2 mm. It is also observed that, in MatVP2DEIT with 2048 elements, the boundary data generated for a phantom with a circular inhomogeneity of a diameter less than 7% of that of the phantom domain can produce resistivity images in EIDORS with a 1968 element mesh. Results also show that the MatVP2DEIT accurately generates the boundary data for neighbouring, opposite reference and trigonometric current patterns which are very suitable for resistivity reconstruction studies. MatVP2DEIT generated data are also found suitable for studying the effect of the different regularization methods on reconstruction process. Comparing the reconstructed image with an original geometry made in MatVP2DEIT, it would be easier to study the resistivity imaging procedures as well as the inverse solver performance. Using the proposed MatVP2DEIT software with modified domains, the cross sectional anatomy of a number of body parts can be simulated in PC and the impedance image reconstruction of human anatomy can be studied.

Generative Maximum Entropy Learning for Multiclass Classification

Maximum entropy approach to classification is very well studied in applied statistics and machine learning and almost all the methods that exists in literature are discriminative in nature. In this paper, we introduce a maximum entropy classification method with feature selection for large dimensional data such as text datasets that is generative in nature. To tackle the curse of dimensionality of large data sets, we employ conditional independence assumption (Naive Bayes) and we perform feature selection simultaneously, by enforcing a `maximum discrimination' between estimated class conditional densities. For two class problems, in the proposed method, we use Jeffreys (J) divergence to discriminate the class conditional densities. To extend our method to the multi-class case, we propose a completely new approach by considering a multi-distribution divergence: we replace Jeffreys divergence by Jensen-Shannon (JS) divergence to discriminate conditional densities of multiple classes. In order to reduce computational complexity, we employ a modified Jensen-Shannon divergence (JS(GM)), based on AM-GM inequality. We show that the resulting divergence is a natural generalization of Jeffreys divergence to a multiple distributions case. As far as the theoretical justifications are concerned we show that when one intends to select the best features in a generative maximum entropy approach, maximum discrimination using J-divergence emerges naturally in binary classification. Performance and comparative study of the proposed algorithms have been demonstrated on large dimensional text and gene expression datasets that show our methods scale up very well with large dimensional datasets.