Biconvex Relaxation for Semidefinite Programming

Total Page:16

File Type:pdf, Size:1020Kb

Biconvex Relaxation for Semidefinite Programming Biconvex Relaxation for Semidefinite Programming in Computer Vision Sohil Shah1, Abhay Yadav1, Tom Goldstein1, David Jacobs1, Carlos Castillo1 and Christoph Studer2 1University of Maryland, College Park, 2Cornell University, Ithaca, NY Abstract non-convex constraints, such as integer-valued vectors and low-rank matrices, can be handled by semidefinite relaxation Semidefinite programming has become an indispensable (SDR) [16]. In this method, a non-convex optimization tool in computer vision. Unfortunately, general-purpose problem involving a vector of unknowns is “lifted” to a solvers for semidefinite programs are too slow to be used on higher dimensional convex problem involving a positive large-scale problems and exhibit excessive memory require- semidefinite matrix (all eigenvalues are non-negative) of ments. We propose a general framework to approximately unknowns, which then enables one to solve a semidefinite solve large-scale semidefinite problems (SDPs) at low com- program (SDP) [31]. While SDR was shown to deliver putational complexity. Our new approach, referred to as state-of-the-art performance in a wide range of applications biconvex relaxation (BCR), transforms a general SDP into [9, 16, 10, 30, 37, 20], the approach significantly increases a specific biconvex optimization problem, which can then be the dimensionality of the original optimization problem (i.e., solved in the original, low-dimensional variable space at replacing a vector with a matrix), which typically results in low complexity. The resulting biconvex problem is approx- exorbitant computational costs and memory requirements. imately solved using an efficient alternating minimization Nevertheless, SDR leads to SDPs whose global optimal (AM) procedure. Since AM has the potential to get stuck in solution(s) can be found using robust numerical methods. local minima, we propose a general initialization scheme A growing number of computer-vision applications in- that enables our BCR framework to start close to the global volve high-resolution images (or even videos) that require optimum of the original SDP—this is key for our algorithm the solution to SDPs with an excessive number of variables. to quickly converge to optimal or near-optimal solutions. We General-purpose solvers for SDPs (e.g., interior point meth- showcase the efficacy of our approach on two applications in ods) require high computational complexity; the worst-case computer vision, namely segmentation and co-segmentation. complexity is O(N 6:5 log(1=")) [25] for an N ×N problem BCR achieves a comparable solution quality to state-of-the- with objective error. Very often, N is proportional to the art SDP methods with speedups ranging between 4× and number of image pixels, which is potentially very large. 35×. At the same time, BCR handles a more general set of SDPs than the fastest previous approaches, which are more The prohibitive complexity and memory requirements of specialized. solving SDPs exactly with a large number of variables has spawned recent interest in fast, non-convex solvers that avoid lifting, i.e., operate in the original low-dimensional variable 1. Introduction space. For example, recent progress in phase retrieval by Candes` et al. [5] has shown that non-convex optimization Optimization problems involving either integer-valued methods provably achieve a solution quality comparable to vectors or matrices with low-rank structure are ubiquitous exact SDR-based methods at significantly lower complexity. in computer vision. Graph-cut methods for image segmen- This method operates on the original dimension, which en- tation, for example, involve optimization problems where ables its use on high-dimensional problems (reference [5] integer-valued variables represent region labels [26, 13, 30, performs phase retrieval on a 1080×1920 pixel image in less 9]. Problems in multi-camera structure-from motion [1], than 22 minutes). Another prominent example is max-norm manifold embedding [37], and matrix completion [20] re- regularization by Lee et al. [15], which was proposed for quire the solution of optimization problems over matrices solving high-dimensional matrix-completion problems and that are constrained to have low-rank. Since the constraints to approximately perform max-cut clustering. This method in such optimization problems are non-convex, the design was shown to outperform exact SDR-based methods in terms of computationally-efficient algorithms that find globally of computational complexity, while delivering acceptable optimal solutions is, in general, a difficult task. solution quality. While both of these examples outperform For a wide range of applications [16, 14, 3, 8, 27, 37], classical SDP-based methods, they are limited to very spe- 1 cific problem types, i.e., they perform one of phase retrieval, vex constraints via semidefinite relaxation (SDR) [16] and matrix completion, or max-cut clustering. More importantly, (ii) the resulting problems often come with strong theoreti- these methods are unable to handle complex SDPs that typi- cal guarantees. For example, the Goemans-Williamson [9] cally appear in computer vision. SDR seeks to find a solution to the NP-complete max-cut problem, which can be relaxed to the form (1) and solved 1.1. Contributions using general-purpose methods. Furthermore, the value of In this paper, we introduce a novel framework for approx- the graph cut obtained via this relaxation is known to exceed imately solving general SDPs in a computationally-efficient 80% of the maximum achievable value. manner and with small memory footprint. Our approach, re- In computer vision, a large number of problems can be ferred to as biconvex relaxation (BCR), transforms a general cast as SDPs of the general form (1). For example, [37] SDP into a specific biconvex optimization problem, which formulates image manifold learning as an SDP, [27] uses can then be solved in the original, low-dimensional vari- an SDP to enforce a non-negative lighting constraint when able space at low complexity. The resulting biconvex prob- recovering scene lighting and object albedos, [2] uses an lem is solved using a computationally-efficient alternating SDP for graph matching, [1] proposes an SDP that recovers minimization (AM) procedure. Since AM is prone to get the orientation of multiple cameras from point correspon- stuck in local minima, we propose a general initialization dences and essential matrices, and [20] uses low-rank SDPs scheme that enables our BCR framework to start close to to solve matrix-completion problems that arise in structure- the global optimum of the original SDP—this initialization from-motion and photometric stereo. is key for our algorithm to quickly converge to an optimal or near-optimal solution. We showcase the effectiveness of 2.2. SDR for Binary-Valued Quadratic Problems the proposed BCR framework by comparing the resulting A large body of results that use SDR to solve binary la- algorithms to that of highly-specialized SDP solvers for a beling problems exist in the literature. For such problems, select set of problems in computer vision involving image a set of variables take on binary values (e.g., indicating a segmentation and co-segmentation. Our results demonstrate segmentation into two groups) while minimizing a quadratic that BCR enables high-quality results at best-in-class run- cost function that depends on the assignment of pairs of vari- times. Specifically, BCR achieves speedups ranging from ables. Such labeling problems typically arise from Markov 4× to 35× over state-of-the-art competitor methods [34, 17] random fields (MRFs) for which a rich literature on solution for image segmentation and co-segmentation. methods exist (see [33] for a recent survey). Spectral meth- ods, e.g., [26], are often used to solve such binary-valued 2. Background and Relevant Prior Art quadratic problems (BQPs)—the references [13, 30] used We now briefly review semidefinite programs (SDPs) and SDR, inspired by the work of [9] that provides a generalized then discuss prior work on fast, approximate solvers for SDR for the max-cut problem. BQP problems have wide SDPs in computer vision and related applications. applicability to computer vision problems, such as segmen- tation and perceptual organization [13, 34, 11], semantic 2.1. Semidefinite Programs (SDPs) segmentation [35], matching [30, 24], surface reconstruction including photometric stereo and shape from defocus [8], SDPs find use in a large (and growing) number of fields, and image restoration [23]. including computer vision, machine learning, signal and im- The key idea of solving BQPs via SDR is to lift the binary- age processing, statistics, communications, and control [31]. N 2 SDPs can be written in the following general form: valued label vector b 2 {±1g to an N -dimensional ma- trix space by forming the PSD matrix B = bbT , whose minimize hC; Yi non-convex rank-1 constraint is relaxed to PSD matrices N×N Y2R + B 2 SN×N with an all-ones diagonal [16]. The goal is then subject to hA ; Yi = b ; 8i 2 E; i i (1) to solve a SDP for B in the hope that the resulting matrix is hAj; Yi ≤ bj; 8j 2 B; rank 1 (in this case, SDR found an optimal solution); if B has Y 2 S+ ; higher rank, an approximate solution must be extracted. Ap- N×N proximate solutions can either be obtained from the leading + eigenvector or via randomization methods [16]. where SN×N represents the set of N × N symmetric, pos- hC; Yi = tr(CT Y) itive semidefinite matrices, and is the 2.3. Specialized Solvers for SDPs matrix inner product. The sets E and B contain the indices associated to the equality and inequality constraints, respec- While a number of general-purpose solvers for SDPs tively; the matrices Ai and Aj have appropriate dimensions. exist, such as SeDuMi [28] or SDPT3 [29], their computa- The key advantages of SDPs are that (i) they enable the tional complexity and memory requirements are generally transformation of certain non-convex constraints into con- high.
Recommended publications
  • Globally Optimal Model-Based Clustering Via Mixed Integer Nonlinear Programming
    Globally Optimal Model-based Clustering via Mixed Integer Nonlinear Programming Patrick Flaherty Pitchaya Wiratchotisatian Department of Mathematics & Statistics Data Science Program UMass Amherst Worcester Polytechnic Institute Amherst, MA 01002 Worcester, MA 01609 [email protected] [email protected] Andrew C. Trapp Foise Business School Worcester Polytechnic Institute Worcester, MA 01609 [email protected] Abstract We present an approach for solving for the optimal partition of a data set via global optimization of a maximum a-posteriori (MAP) estimation problem. Our approach formulates the MAP estimation problem for the Gaussian mixture model as a mixed-integer nonlinear optimization problem (MINLP). Our method provides a certificate of global optimality, can accommodate side constraints, and is extendable to other finite mixture models. We propose an approximation to the MINLP that transforms it into a mixed integer quadratic program (MIQP) which preserves global optimality within desired accuracy and improves computational aspects. Numerical experiments compare our method to standard estimation approaches and show that our method finds the globally optimal MAP for some standard data sets, providing a benchmark for comparing estimation methods. 1 Introduction In the application of clustering models to real data there is often rich prior information about the relationships among the samples or the relationships between the samples and the parameters. For example, in biological or clinical experiments, it may be known that two samples are technical replicates and should belong to the same cluster, or it may be known that the mean value for certain control samples is in a certain range. However, standard model-based clustering methods make it difficult to enforce such hard logical constraints.
    [Show full text]
  • Attenuation Imaging by Wavefield Reconstruction Inversion with Bound
    Attenuation imaging by wavefield reconstruction inversion with bound constraints and total variation regularization Hossein S. Aghamiry 12, Ali Gholami 1 and St´ephaneOperto 2 INTRODUCTION ABSTRACT Wavefield reconstruction inversion (WRI) extends the Full waveform inversion (FWI) is a high-resolution non- search space of Full Waveform Inversion (FWI) by al- linear imaging technology which can provide accurate sub- lowing for wave equation errors during wavefield re- surface model by matching observed and calculated wave- construction to match the data from the first iteration. forms (Tarantola, 1984; Pratt et al., 1998; Virieux and Then, the wavespeeds are updated from the wavefields Operto, 2009). However, it is well acknowledged that it by minimizing the source residuals. Performing these suffers from two main pathologies. The first one is the two tasks in alternating mode breaks down the non- nonlinearity associated with cycle skipping: When the linear FWI as a sequence of two linear subproblems, distance between the observed and calculated data is the relaying on the bilinearity of the wave equation. We least-squares norm of their differences, FWI remains stuck solve this biconvex optimization with the alternating- into spurious local minima when the initial velocity model direction method of multipliers (ADMM) to cancel out does not allow to match traveltimes with an error lower efficiently the data and source residuals in iterations than half a period. To mitigate cycle skipping, many vari- and stabilize the parameter estimation with appropri- ants of FWI have been proposed with more convex dis- ate regularizations. Here, we extend WRI to viscoa- tances such as those based on matching filters (Warner coustic media for attenuation imaging.
    [Show full text]
  • Arxiv:1608.04430V3 [Math.OC] 28 Dec 2018 R R 0 Counts the Number of Nonzero Elements in a Vector
    Sparsity Constrained Minimization via Mathematical Programming with Equilibrium Constraints Ganzhao Yuan [email protected] School of Data & Computer Science, Sun Yat-sen University (SYSU), P.R. China Bernard Ghanem [email protected] King Abdullah University of Science & Technology (KAUST), Saudi Arabia Editor: Abstract Sparsity constrained minimization captures a wide spectrum of applications in both machine learning and signal processing. This class of problems is difficult to solve since it is NP-hard and existing solutions are primarily based on Iterative Hard Thresholding (IHT). In this paper, we consider a class of continuous optimization techniques based on Mathematical Programs with Equilibrium Constraints (MPECs) to solve general sparsity constrained problems. Specifically, we reformulate the problem as an equivalent biconvex MPEC, which we can solve using an exact penalty method or an alternating direction method. We elaborate on the merits of both proposed methods and analyze their convergence properties. Finally, we demonstrate the effectiveness and versatility of our methods on several important problems, including feature selection, segmented regression, trend filtering, MRF optimization and impulse noise removal. Extensive experiments show that our MPEC-based methods outperform state-of-the-art techniques, especially those based on IHT. Keywords: Sparsity Constrained Optimization, Exact Penalty Method, Alternating Direction Method, MPEC, Convergence Analysis 1. Introduction In this paper, we mainly focus on the following generalized sparsity constrained minimization prob- lem: min f(x); s:t: kAxk0 ≤ k (1) x where x 2 n, A 2 m×n and k (0 < k < m) is a positive integer. k · k is a function that arXiv:1608.04430v3 [math.OC] 28 Dec 2018 R R 0 counts the number of nonzero elements in a vector.
    [Show full text]
  • Cosparse Regularization of Physics-Driven Inverse Problems Srdan Kitic
    Cosparse regularization of physics-driven inverse problems Srdan Kitic To cite this version: Srdan Kitic. Cosparse regularization of physics-driven inverse problems. Signal and Image Processing. Université Rennes 1, 2015. English. NNT : 2015REN1S152. tel-01237323v3 HAL Id: tel-01237323 https://hal.archives-ouvertes.fr/tel-01237323v3 Submitted on 19 Aug 2016 HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non, lished or not. The documents may come from émanant des établissements d’enseignement et de teaching and research institutions in France or recherche français ou étrangers, des laboratoires abroad, or from public or private research centers. publics ou privés. ANNÉE 2015 THÈSE / UNIVERSITÉ DE RENNES 1 sous le sceau de l’Université Européenne de Bretagne pour le grade de DOCTEUR DE L’UNIVERSITÉ DE RENNES 1 Mention : Traitement du Signal et Télécommunications Ecole doctorale MATISSE présentée par Srđan Kitić préparée à l’unité de recherche 6074 IRISA (Institut de Recherche en Informatique et Systèmes Aléatoires) Thèse soutenue à l’IRISA le 26 novembre 2015 devant le jury composé de : Cosparse regularization Laurent DAUDET Professeur, Université Paris 7, Laboratoire Ondes et of physics-driven Acoustiques / rapporteur Patrick FLANDRIN Directeur de recherche CNRS, ENS de Lyon, inverse problems Membre de l’Académie des Sciences / rapporteur Michael DAVIES Professeur, University of Edinburgh / examinateur Laurent DUVAL Ingénieur de Recherche, chef de projet, IFP Energies Nouvelles / examinateur Laurent ALBERA Maître de conférences, Université de Rennes 1, LTSI / membre invité Rémi GRIBONVAL Directeur de recherche INRIA / directeur de thèse Nancy BERTIN Chargé de recherche CNRS / co-directrice de thèse Dedicated to my family Acknowledgements I am deeply thankful to my supervisors for their guidance during the last three years.
    [Show full text]
  • A Primal Method for Multiple Kernel Learning
    Neural Comput & Applic DOI 10.1007/s00521-012-1022-2 ORIGINAL ARTICLE A primal method for multiple kernel learning Zhifeng Hao • Ganzhao Yuan • Xiaowei Yang • Zijie Chen Received: 12 February 2012 / Accepted: 21 June 2012 Ó Springer-Verlag London Limited 2012 Abstract The canonical support vector machines (SVMs) direction can be computed using several Preconditioned are based on a single kernel, recent publications have shown Conjugate Gradient steps on the Hessian operator equation, that using multiple kernels instead of a single one can the algorithm is shown more efficient than the current primal enhance interpretability of the decision function and promote approaches in this MKL setting. Furthermore, we use the classification accuracy. However, most of existing approa- Nesterov’s optimal gradient method to optimize the kernel ches mainly reformulate the multiple kernel learning as a weights. One remarkable advantage of solving in the primal is saddle point optimization problem which concentrates on that it achieves much faster convergence rate than solving in solving the dual. In this paper, we show that the multiple the dual and does not require a two-stage algorithm even for kernel learning (MKL) problem can be reformulated as a the single kernel LapSVM. Introducing the Laplacian regu- BiConvex optimization and can also be solved in the primal. larizer, we also extend our primal method to semi-supervised While the saddle point method still lacks convergence results, scenario. Extensive experiments on some UCI benchmarks our proposed method exhibits strong optimization conver- have shown that the proposed algorithm converges rapidly gence properties. To solve the MKL problem, a two-stage and achieves competitive accuracy.
    [Show full text]
  • Binary Optimization Via Mathematical Programming with Equilibrium Constraints
    1 Binary Optimization via Mathematical Programming with Equilibrium Constraints Ganzhao Yuan, Bernard Ghanem Abstract—Binary optimization is a central problem in mathematical optimization and its applications are abundant. To solve this problem, we propose a new class of continuous optimization techniques which is based on Mathematical Programming with Equilibrium Constraints (MPECs). We first reformulate the binary program as an equivalent augmented biconvex optimization problem with a bilinear equality constraint, then we propose two penalization/regularization methods (exact penalty and alternating direction) to solve it. The resulting algorithms seek desirable solutions to the original problem via solving a sequence of linear programming convex relaxation subproblems. In addition, we prove that both the penalty function and augmented Lagrangian function, induced by adding the complementarity constraint to the objectives, are exact, i.e., they have the same local and global minima with those of the original binary program when the penalty parameter is over some threshold. The convergence of both algorithms can be guaranteed, since they essentially reduce to block coordinate descent in the literature. Finally, we demonstrate the effectiveness and versatility of our methods on several important problems, including graph bisection, constrained image segmentation, dense subgraph discovery, modularity clustering and Markov random fields. Extensive experiments show that our methods outperform existing popular techniques, such as iterative hard thresholding, linear programming relaxation and semidefinite programming relaxation. Index Terms—Binary Optimization, Convergence Analysis, MPECs, Exact Penalty Method, Alternating Direction Method, Graph Bisection, Constrained Image Segmentation, Dense Subgraph Discovery, Modularity Clustering, Markov Random Fields. F 1 INTRODUCTION Another type of optimization focuses on the cutting-plane and In this paper, we mainly focus on the following binary optimiza- branch-and-cut method.
    [Show full text]
  • New SOCP Relaxation and Branching Rule for Bipartite Bilinear Programs
    New SOCP relaxation and branching rule for bipartite bilinear programs Santanu S. Dey∗1, Asteroide Santanay1, and Yang Wangz2 1School of Industrial and Systems Engineering, Georgia Institute of Technology 2School of Civil and Environmental Engineering, Georgia Institute of Technology March 26, 2018 Abstract A bipartite bilinear program (BBP) is a quadratically constrained quadratic optimiza- tion problem where the variables can be partitioned into two sets such that fixing the variables in any one of the sets results in a linear program. We propose a new second order cone representable (SOCP) relaxation for BBP, which we show is stronger than the stan- dard SDP relaxation intersected with the boolean quadratic polytope. We then propose a new branching rule inspired by the construction of the SOCP relaxation. We describe a new application of BBP called as the finite element model updating problem, which is a fundamental problem in structural engineering. Our computational experiments on this problem class show that the new branching rule together with an polyhedral outer approx- imation of the SOCP relaxation outperforms a state-of-the-art commercial global solver in obtaining dual bounds. 1 Introduction: Bipartite bilinear program (BBP) A quadratically constrained quadratic program (QCQP) is called as a bilinear optimization problem if every degree two term in the constraints and objective involves the product of two distinct variables. For a given instance of bilinear optimization problem, one often associates a simple graph constructed as follows: The set of vertices corresponds to the variables in the instance and there is an edge between two vertices if there is a degree two term involving the corresponding variables in the instance formulation.
    [Show full text]
  • New SOCP Relaxation and Branching Rule for Bipartite Bilinear Programs
    New SOCP relaxation and branching rule for bipartite bilinear programs Santanu S. Dey∗1, Asteroide Santanay1, and Yang Wangz2 1School of Industrial and Systems Engineering, Georgia Institute of Technology 2School of Civil and Environmental Engineering, Georgia Institute of Technology March 25, 2018 Abstract A bipartite bilinear program (BBP) is a quadratically constrained quadratic optimiza- tion problem where the variables can be partitioned into two sets such that fixing the variables in any one of the sets results in a linear program. We propose a new second order cone representable (SOCP) relaxation for BBP, which we show is stronger than the stan- dard SDP relaxation intersected with the boolean quadratic polytope. We then propose a new branching rule inspired by the construction of the SOCP relaxation. We describe a new application of BBP called as the finite element model updating problem, which is a fundamental problem in structural engineering. Our computational experiments on this problem class show that the new branching rule together with an polyhedral outer approx- imation of the SOCP relaxation outperforms a state-of-the-art commercial global solver in obtaining dual bounds. 1 Introduction: Bipartite bilinear program (BBP) A quadratically constrained quadratic program (QCQP) is called as a bilinear optimization problem if every degree two term in the constraints and objective involves the product of two distinct variables. For a given instance of bilinear optimization problem, one often associates a simple graph constructed as follows: The set of vertices corresponds to the variables in the instance and there is an edge between two vertices if there is a degree two term involving the corresponding variables in the instance formulation.
    [Show full text]
  • Minimum Spanning Trees with Neighborhoods Let G = (V, E) Be a Connected Undirected Graph, Whose Vertices Are Embedded D D D in R , I.E., V ∈ R for All V ∈ V
    MINIMUM SPANNING TREES WITH NEIGHBORHOODS V´ICTOR BLANCO, ELENA FERNANDEZ,´ AND JUSTO PUERTO Abstract. This paper studies Minimum Spanning Trees under incomplete information for its vertices. We assume that no information is available on the precise placement of vertices so that it is only known that vertices belong to some neighborhoods that are second order cone representable and distances are measured with a ℓq-norm. Two mixed integer non linear mathematical programming formulations are presented, based on alternative representations of subtour elimination constraints. A solution scheme is also proposed, result- ing from a reformulation suitable for a Benders-like decomposition, which is embedded within an exact branch-and-cut framework. Furthermore, a math- heuristic is developed, which alternates in solving convex subproblems in dif- ferent solution spaces, and is able to solve larger instances. The results of extensive computational experiments are reported and analyzed. 1. Introduction Nowadays Combinatorial Optimization (CO) lies in the heart of multiple appli- cations in the field of Operations Research. Many such applications can be formu- lated as optimization problems defined on graphs where some particular structure is sought satisfying some optimality property. Traditionally this type of problems assumed implicitly the exact knowledge of all input elements, and, in particular, of the precise position of vertices and edges. Nevertheless, this assumption does not always hold, as uncertainty, lack of information, or some other factors may affect the relative position of the elements of the input graph. Hence, new tools are required to give adequate answers to these challenges, which have been often ignored by standard CO tools.
    [Show full text]
  • Chemical and Phase Equilibria Through Deterministic Global Optimization
    UNIVERSIDADE FEDERAL DE MINAS GERAIS MÁRCIO RIBEIRO VIANNA NETO CHEMICAL AND PHASE EQUILIBRIA THROUGH DETERMINISTIC GLOBAL OPTIMIZATION BELO HORIZONTE 2016 MÁRCIO RIBEIRO VIANNA NETO CHEMICAL AND PHASE EQUILIBRIA THROUGH DETERMINISTIC GLOBAL OPTIMIZATION Dissertação apresentada ao Curso de Pós-Graduação em Engenharia Química como parte dos requisitos exigidos para a obtenção do título de MESTRE EM ENGENHARIA QUÍMICA. Orientador: Prof. Dr. Éder Domingos de Oliveira Belo Horizonte 2016 ii To my parents, Marcos Vianna and Rita Vianna, to whom I owe all that is best in me. iii ACKNOWLEDGEMENTS I would like to thank my parents, Marcos and Rita, for all their love, patience and never-ending support. My good friend Bernardo who has always been there and supported all my endeavors. Also, my thanks to Professor Éder Domingos for his continuous support, for all his trust, for all the freedom that he has given me during my studies, and lastly, for his seemingly endless patience. My thanks to Professor Marcelo Cardoso for his friendship, for all the opportunities with which he has provided me and for his invaluable advice. My thanks to all my colleagues at the Industrial Processes Laboratory for being such an engaging, welcoming and supportive team. My dearest thanks to all the people who have indirectly contributed to this work, either by giving me bits of advice or by keeping me sane. My many thanks to Professor Ívina Paula for being my first chemistry teacher, and also Professor Miriam Miranda for encouraging me to do research. Lastly, I would like to thank CNPq for their financial support.
    [Show full text]
  • Parametric Dual Maximization for Non-Convex Learning Problems
    Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence (AAAI-17) Parametric Dual Maximization for Non-Convex Learning Problems Yuxun Zhou Zhaoyi Kang Costas J. Spanos Department of EECS Department of EECS Department of EECS UC Berkeley UC Berkeley UC Berkeley [email protected] [email protected] [email protected] Abstract solving standard SVM and updating latent variables. An- other widely applied technique is the concave-convex Proce- We consider a class of non-convex learning problems that can dure (CCCP) (Yuille, Rangarajan, and Yuille 2002). Among be formulated as jointly optimizing regularized hinge loss and a set of auxiliary variables. Such problems encompass but are many others, (Yu and Joachims 2009; Ping, Liu, and Ihler not limited to various versions of semi-supervised learning, 2014) used CCCP for latent structural SVM training. Direct learning with hidden structures, robust learning, etc. Existing application of the gradient-based method is especially attrac- methods either suffer from local minima or have to invoke a tive for large scale problems owing to its low computational non-scalable combinatorial search. In this paper, we propose cost (Bottou 2010). Such examples include the stochastic a novel learning procedure, namely Parametric Dual Max- gradient descent (SGD) for large margin polytope machine imization (PDM), that can approach global optimality effi- (Kantchelian et al. 2014; Zhou, Jin, and Spanos 2015) and ciently with user specified approximation levels. The building S3VM (Chapelle and Zien 2005). Combinatorial optimiza- blocks of PDM are two new results: (1) The equivalent con- tion methods, e.g., the local search method (Joachims 1999) vex maximization reformulation derived by parametric anal- and branch and bound (B & B) (Chapelle, Sindhwani, and ysis.
    [Show full text]
  • Biconvex Relaxation for Semidefinite Programming in Computer Vision
    Biconvex Relaxation for Semidefinite Programming in Computer Vision Sohil Shah1(B),AbhayKumarYadav1,CarlosD.Castillo1,DavidW.Jacobs1, Christoph Studer2,andTomGoldstein1 1 University of Maryland, College Park, MD, USA [email protected], jaiabhay,djacobs,tomg @cs.umd.edu, [email protected]{ } 2 Cornell University, Ithaca, NY, USA [email protected] Abstract. Semidefinite programming (SDP) is an indispensable tool in computer vision, but general-purpose solvers for SDPs are often too slow and memory intensive for large-scale problems. Our framework, referred to as biconvex relaxation (BCR), transforms an SDP consisting of PSD constraint matrices into a specific biconvex optimization problem, which can then be approximately solved in the original, low-dimensional vari- able space at low complexity. The resulting problem is solved using an efficient alternating minimization (AM) procedure. Since AM has the potential to get stuck in local minima, we propose a general initializa- tion scheme that enables BCR to start close to a global optimum—this is key for BCR to quickly converge to optimal or near-optimal solu- tions. We showcase the efficacy of our approach on three applications in computer vision, namely segmentation, co-segmentation, and manifold metric learning. BCR achieves solution quality comparable to state-of- the-art SDP methods with speedups between 4 and 35 . × × 1Introduction Optimization problems involving either integer-valued vectors or low-rank matri- ces are ubiquitous in computer vision. Graph-cut methods for image segmenta- tion, for example, involve optimization problems where integer-valued variables represent region labels [1–4]. Problems in multi-camera structure from motion [5], manifold embedding [6], and matrix completion [7]allrelyonoptimization problems involving matrices with low rank constraints.
    [Show full text]