Slack and Margin Rescaling as Convex Extensions of Supermodular Functions

Slack and margin rescaling are variants of the structured output SVM, which is frequently applied to problems in computer vision such as image segmentation, object localization, and learning parts based object models. They define convex surrogates to task specific loss functions, which, when specialized to non-additive loss functions for multi-label problems, yield extensions to increasing set functions. We demonstrate in this paper that we may use these concepts to define polynomial time convex extensions of arbitrary supermodular functions, providing an analysis framework for the tightness of these surrogates. This analysis framework shows that, while neither margin nor slack rescaling dominate the other, known bounds on supermodular functions can be used to derive extensions that dominate both of these, indicating possible directions for defining novel structured output prediction surrogates. In addition to the analysis of structured prediction loss functions, these results imply an approach to supermodular minimization in which margin rescaling is combined with non-polynomial time convex extensions to compute a sequence of LP relaxations reminiscent of a cutting plane method. This approach is applied to the problem of selecting representative exemplars from a set of images, validating our theoretical contributions.

[1]  Andreas Krause,et al.  Submodular Function Maximization , 2014, Tractability.

[2]  Gérard Cornuéjols,et al.  Integer programming , 2014, Math. Program..

[3]  Jan Vondr Submodularity and Curvature: The Optimal Algorithm , 2010 .

[4]  Satoru Fujishige,et al.  Submodular functions and optimization , 1991 .

[5]  Gökhan BakIr,et al.  Generalization Bounds and Consistency for Structured Labeling , 2007 .

[6]  Jeff A. Bilmes,et al.  Submodularity beyond submodular energies: Coupling edges in graph cuts , 2011, CVPR 2011.

[7]  Thomas Hofmann,et al.  Large Margin Methods for Structured and Interdependent Output Variables , 2005, J. Mach. Learn. Res..

[8]  Jan Vondrák,et al.  Optimal approximation for the submodular welfare problem in the value oracle model , 2008, STOC.

[9]  Rishabh K. Iyer,et al.  Algorithms for Approximate Minimization of the Difference Between Submodular Functions, with Applications , 2012, UAI.

[10]  Rishabh K. Iyer,et al.  Polyhedral aspects of Submodularity, Convexity and Concavity , 2015, ArXiv.

[11]  Matthew B. Blaschko,et al.  A Convex Surrogate Operator for General Non-Modular Loss Functions , 2016, AISTATS.

[12]  Adrian Weller,et al.  Tightness of LP Relaxations for Almost Balanced Models , 2016, AISTATS.

[13]  Joseph Naor,et al.  A Tight Linear Time (1/2)-Approximation for Unconstrained Submodular Maximization , 2012, 2012 IEEE 53rd Annual Symposium on Foundations of Computer Science.

[14]  Vladimir Kolmogorov,et al.  Analyzing Convex Relaxations for MAP Estimation , 2009 .

[15]  László Lovász,et al.  Submodular functions and convexity , 1982, ISMP.

[16]  M. L. Fisher,et al.  An analysis of approximations for maximizing submodular set functions—I , 1978, Math. Program..

[17]  Ofer Meshi,et al.  Fast and Scalable Structural SVM with Slack Rescaling , 2015, AISTATS.

[18]  Fredrik Kahl,et al.  Generalized roof duality , 2012, Discret. Appl. Math..

[19]  Rishabh K. Iyer,et al.  Curvature and Optimal Algorithms for Learning and Minimizing Submodular Functions , 2013, NIPS.

[20]  Gérard Cornuéjols,et al.  Submodular set functions, matroids and the greedy algorithm: Tight worst-case bounds and some generalizations of the Rado-Edmonds theorem , 1984, Discret. Appl. Math..

[21]  Antonio Torralba,et al.  Ieee Transactions on Pattern Analysis and Machine Intelligence 1 80 Million Tiny Images: a Large Dataset for Non-parametric Object and Scene Recognition , 2022 .

[22]  Matthew B. Blaschko,et al.  Learning Submodular Losses with the Lovasz Hinge , 2015, ICML.