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.

Further results on geometric properties of a family of relative entropies

This paper extends some geometric properties of a one-parameter family of relative entropies. These arise as redundancies when cumulants of compressed lengths are considered instead of expected compressed lengths. These parametric relative entropies are a generalization of the Kullback-Leibler divergence. They satisfy the Pythagorean property and behave like squared distances. This property, which was known for finite alphabet spaces, is now extended for general measure spaces. Existence of projections onto convex and certain closed sets is also established. Our results may have applications in the Rényi entropy maximization rule of statistical physics.

Geometry of static balancing during posture transition

The interaction between the digital human model (DHM) and environment typically occurs in two distinct modes; one, when the DHM maintains contacts with the environment using its self weight, wherein associated reaction forces at the interface due to gravity are unidirectional; two, when the DHM applies both tension and compression on the environment through anchoring. For static balancing in first mode of interaction, it is sufficient to maintain the projection of the centre of mass (COM) inside the convex region induced by the weight supporting segments of the body on a horizontal plane. In DHM, static balancing is required while performing specified tasks such as reach, manipulation and locomotion; otherwise the simulations would not be realistic. This paper establishes the geometric relationships that must be satisfied for maintaining static balance while altering the support configurations for a given posture and altering the posture for a given support condition. For a given location of the COM for a system supported by multiple point contacts, the conditions for simultaneous withdrawal of a specified set of contacts have been determined in terms of the convex hulls of the subsets of the points of contact. When the projection of COM must move beyond the existing support for performing some task, new supports must be enabled for maintaining static balance. This support seeking behavior could also manifest while planning for reduction of support stresses. Feasibility of such a support depends upon the availability of necessary features in the environment. Geometric conditions necessary for selection of new support on horizontal,inclined and vertical surfaces within the workspace of the DHM for such dynamic scenario have been derived. The concepts developed are demonstrated using the cases of sit-to-stand posture transition for manipulation of COM within the convex supporting polygon, and statically stable walking gaits for support seeking within the kinematic capabilities of the DHM. The theory developed helps in making the DHM realize appropriate behaviors in diverse scenarios autonomously.

IceCube: efficient targeted mining in data cubes

We address the problem of mining targeted association rules over multidimensional market-basket data. Here, each transaction has, in addition to the set of purchased items, ancillary dimension attributes associated with it. Based on these dimensions, transactions can be visualized as distributed over cells of an n-dimensional cube. In this framework, a targeted association rule is of the form {X -> Y} R, where R is a convex region in the cube and X. Y is a traditional association rule within region R. We first describe the TOARM algorithm, based on classical techniques, for identifying targeted association rules. Then, we discuss the concepts of bottom-up aggregation and cubing, leading to the CellUnion technique. This approach is further extended, using notions of cube-count interleaving and credit-based pruning, to derive the IceCube algorithm. Our experiments demonstrate that IceCube consistently provides the best execution time performance, especially for large and complex data cubes.

Robust formulations for clustering-based large-scale classification

Chebyshev-inequality-based convex relaxations of Chance-Constrained Programs (CCPs) are shown to be useful for learning classifiers on massive datasets. In particular, an algorithm that integrates efficient clustering procedures and CCP approaches for computing classifiers on large datasets is proposed. The key idea is to identify high density regions or clusters from individual class conditional densities and then use a CCP formulation to learn a classifier on the clusters. The CCP formulation ensures that most of the data points in a cluster are correctly classified by employing a Chebyshev-inequality-based convex relaxation. This relaxation is heavily dependent on the second-order statistics. However, this formulation and in general such relaxations that depend on the second-order moments are susceptible to moment estimation errors. One of the contributions of the paper is to propose several formulations that are robust to such errors. In particular a generic way of making such formulations robust to moment estimation errors is illustrated using two novel confidence sets. An important contribution is to show that when either of the confidence sets is employed, for the special case of a spherical normal distribution of clusters, the robust variant of the formulation can be posed as a second-order cone program. Empirical results show that the robust formulations achieve accuracies comparable to that with true moments, even when moment estimates are erroneous. Results also illustrate the benefits of employing the proposed methodology for robust classification of large-scale datasets.

Exchange-spring mechanism of soft and hard ferrite nanocomposites

The paper reports exchange-spring soft and hard ferrite nanocomposites synthesized by chemical co-precipitation with or without the application of ultrasonic vibration. The composites contained BaFe12O19 as the hard phase and CoFe2O4/MgFe2O4 as the soft phase. X-ray diffraction patterns of the samples in the optimum calcined condition indicated the presence of soft ferrites as face-centred cubic (fcc) and hard ferrites as hexagonal close packed (hcp) structure respectively. Temperature dependence of magnetization in the range of 20-700 degrees C demonstrated distinct presence of soft and hard ferrites as magnetic phases which are characterized by wide difference in magnetic anisotropy and coercivity. Exchange-spring mechanism led these nanocomposite systems to exchange-coupled, which ultimately produced convex hysteresis loops characteristic of a single-phase permanent magnet. Fairly high value of coercivity and maximum energy product were observed for the samples in the optimum calcined conditions with a maximum applied field of 1600 kA/m (2 T).

Bounds on the sum-rate capacity of the Gaussian MIMO X channel

We consider the MIMO X channel (XC), a system consisting of two transmit-receive pairs, where each transmitter communicates with both the receivers. Both the transmitters and receivers are equipped with multiple antennas. First, we derive an upper bound on the sum-rate capacity of the MIMO XC under individual power constraint at each transmitter. The sum-rate capacity of the two-user multiple access channel (MAC) that results when receiver cooperation is assumed forms an upper bound on the sum-rate capacity of the MIMO XC. We tighten this bound by considering noise correlation between the receivers and deriving the worst noise covariance matrix. It is shown that the worst noise covariance matrix is a saddle-point of a zero-sum, two-player convex-concave game, which is solved through a primal-dual interior point method that solves the maximization and the minimization parts of the problem simultaneously. Next, we propose an achievable scheme which employs dirty paper coding at the transmitters and successive decoding at the receivers. We show that the derived upper bound is close to the achievable region of the proposed scheme at low to medium SNRs.

Decode-and-forward relay beamforming for secrecy with imperfect CSI and multiple eavesdroppers

In this paper, we evaluate secrecy rates in cooperative relay beamforming in the presence of imperfect channel state information (CSI) and multiple eavesdroppers. A source-destination pair aided by.. out of.. relays, 1 <= k <= M, using decode-and-forward relay beamforming is considered. We compute the worst case secrecy rate with imperfect CSI in the presence of multiple eavesdroppers, where the number of eavesdroppers can be more than the number of relays. We solve the optimization problem for all possible relay combinations to find the secrecy rate and optimum source and relay weights subject to a total power constraint. We relax the rank-1 constraint on the complex semi-definite relay weight matrix and use S-procedure to reformulate the optimization problem that can be solved using convex semi-definite programming.

Scalar transport in diffusion flame wrapped up by an air and fuel side vortex

The present work involves a computational study of soot (chosen as a scalar which is a primary pollutant source) formation and transport in a laminar acetylene diffusion flame perturbed by a convecting line vortex. The topology of soot contours resulting from flame vortex interactions has been investigated. More soot was produced when vortex was introduced from the air side in comparison to the fuel side. Also, the soot topography was spatially more diffuse in the case of air side vortex. The computational model was found to be in good agreement with the experimental work previously reported in the literature. The computational simulation enabled a study of various parameters like temperature, equivalence ratio and temperature gradient affecting the soot production and transport. Temperatures were found to be higher in the case of air side vortex in contrast to the fuel side one. In case of fuel side vortex, abundance of fuel in the vortex core resulted in fuel-rich combustion zone in the core and a more discrete soot topography. Besides, the overall soot production was observed to be low in the fuel side vortex. However, for the air side vortex, air abundance in the core resulted in higher temperatures and greater soot production. Probability density functions (PDFs) have been introduced to investigate the spatiotemporal variation of soot yield and transport and their dependence on temperature and acetylene concentration from statistical view point. In addition, the effect of flame curvature on soot production is also studied. The regions convex to fuel stream side witnessed thicker soot layer. All numerical simulations have been carried out on Fluent 6.3.26. (C) 2013 Elsevier Ltd. All rights reserved.

On the Erdos-Szekeres n-interior-point problem

The n-interior-point variant of the Erdos Szekeres problem is the following: for every n, n >= 1, does there exist a g(n) such that every point set in the plane with at least g(n) interior points has a convex polygon containing exactly n interior points. The existence of g(n) has been proved only for n <= 3. In this paper, we show that for any fixed r >= 2, and for every n >= 5, every point set having sufficiently large number of interior points and at most r convex layers contains a subset with exactly n interior points. We also consider a relaxation of the notion of convex polygons and show that for every n, n >= 1, any point set with at least n interior points has an almost convex polygon (a simple polygon with at most one concave vertex) that contains exactly n interior points. (C) 2013 Elsevier Ltd. All rights reserved.

On piercing (pseudo)lines and boxes

We say a family of geometric objects C has (l;k)-property if every subfamily C0C of cardinality at most lisk- piercable. In this paper we investigate the existence of g(k;d)such that if any family of objects C in Rd has the (g(k;d);k)-property, then C is k-piercable. Danzer and Gr̈ unbaum showed that g(k;d)is infinite for fami-lies of boxes and translates of centrally symmetric convex hexagons. In this paper we show that any family of pseudo-lines(lines) with (k2+k+ 1;k)-property is k-piercable and extend this result to certain families of objects with discrete intersections. This is the first positive result for arbitrary k for a general family of objects. We also pose a relaxed ver-sion of the above question and show that any family of boxes in Rd with (k2d;k)-property is 2dk- piercable.

Classification calibration dimension for general multiclass losses

We study consistency properties of surrogate loss functions for general multiclass classification problems, defined by a general loss matrix. We extend the notion of classification calibration, which has been studied for binary and multiclass 0-1 classification problems (and for certain other specific learning problems), to the general multiclass setting, and derive necessary and sufficient conditions for a surrogate loss to be classification calibrated with respect to a loss matrix in this setting. We then introduce the notion of \emph{classification calibration dimension} of a multiclass loss matrix, which measures the smallest `size' of a prediction space for which it is possible to design a convex surrogate that is classification calibrated with respect to the loss matrix. We derive both upper and lower bounds on this quantity, and use these results to analyze various loss matrices. In particular, as one application, we provide a different route from the recent result of Duchi et al.\ (2010) for analyzing the difficulty of designing `low-dimensional' convex surrogates that are consistent with respect to pairwise subset ranking losses. We anticipate the classification calibration dimension may prove to be a useful tool in the study and design of surrogate losses for general multiclass learning problems.

Curvature based mobility analysis and form closure of smooth planar curves with multiple contacts

This paper presents a simple second-order, curvature based mobility analysis of planar curves in contact. The underlying theory deals with penetration and separation of curves with multiple contacts, based on relative configuration of osculating circles at points of contact for a second-order rotation about each point of the plane. Geometric and analytical treatment of mobility analysis is presented for generic as well as special contact geometries. For objects with a single contact, partitioning of the plane into four types of mobility regions has been shown. Using point based composition operations based on dual-number matrices, analysis has been extended to computationally handle multiple contacts scenario. A novel color coded directed line has been proposed to capture the contact scenario. Multiple contacts mobility is obtained through intersection of the mobility half-spaces. It is derived that mobility region comprises a pair of unbounded or a single bounded convex polygon. The theory has been used for analysis and synthesis of form closure configurations, revolute and prismatic kinematic pairs. (C) 2013 Elsevier Ltd. All rights reserved.

Scalable sequential alternating proximal methods for sparse structural SVMs and CRFs

Structural Support Vector Machines (SSVMs) and Conditional Random Fields (CRFs) are popular discriminative methods used for classifying structured and complex objects like parse trees, image segments and part-of-speech tags. The datasets involved are very large dimensional, and the models designed using typical training algorithms for SSVMs and CRFs are non-sparse. This non-sparse nature of models results in slow inference. Thus, there is a need to devise new algorithms for sparse SSVM and CRF classifier design. Use of elastic net and L1-regularizer has already been explored for solving primal CRF and SSVM problems, respectively, to design sparse classifiers. In this work, we focus on dual elastic net regularized SSVM and CRF. By exploiting the weakly coupled structure of these convex programming problems, we propose a new sequential alternating proximal (SAP) algorithm to solve these dual problems. This algorithm works by sequentially visiting each training set example and solving a simple subproblem restricted to a small subset of variables associated with that example. Numerical experiments on various benchmark sequence labeling datasets demonstrate that the proposed algorithm scales well. Further, the classifiers designed are sparser than those designed by solving the respective primal problems and demonstrate comparable generalization performance. Thus, the proposed SAP algorithm is a useful alternative for sparse SSVM and CRF classifier design.

Sensitivity of Water Dynamics to Biologically Significant Surfaces of Monomeric Insulin: Role of Topology and Electrostatic Interactions

In addition to the biologically active monomer of the protein insulin circulating in human blood, the molecule also exists in dimeric and hexameric forms that are used as storage. The insulin monomer contains two distinct surfaces, namely, the dimer forming surface (DFS) and the hexamer forming surface (HFS), that are specifically designed to facilitate the formation of the dimer and the hexamer, respectively. In order to characterize the structural and dynamical behavior of interfacial water molecules near these two surfaces (DFS and HFS), we performed atomistic molecular dynamics simulations of insulin with explicit water. Dynamical characterization reveals that the structural relaxation of the hydrogen bonds formed between the residues of DFS and the interfacial water molecules is faster than those formed between water and that of the HFS. Furthermore, the residence times of water molecules in the protein hydration layer for both the DFS and HFS are found to be significantly higher than those for some of the other proteins studied so far, such as HP-36 and lysozyme. In particular, we find that more structured water molecules, with higher residence times (similar to 300-500 ps), are present near HFS than those near DFS. A significant slowing down is observed in the decay of associated rotational auto time correlation functions of O-H bond vector of water in the vicinity of HFS. The surface topography and the arrangement of amino acid residues work together to organize the water molecules in the hydration layer in order to provide them with a preferred orientation. HFS having a large polar solvent accessible surface area and a convex extensive nonpolar region, drives the surrounding water molecules to acquire predominantly an outward H-atoms directed, clathrate-like structure. In contrast, near the DFS, the surrounding water molecules acquire an inward H-atoms directed orientation owing to the flat curvature of hydrophobic surface and the interrupted hydrophilic residual alignment. We have followed escape trajectory of several such quasi-bound water molecules from both the surfaces that reveal the significant differences between the two hydration layers.