PASCAL - Pattern Analysis, Statistical Modelling and Computational Learning

An indefinite quadratic formulation for sparse vector optimization
Dori Peleg and Ron Meir
Signal Processing 2006.

Abstract

Sparsity plays an important role in many fields of engineering. The cardinality penalty function, aka the “zero norm”, is neither continuous nor differentiable and therefore smooth optimization algorithms cannot be applied directly. In this paper we present a continuous yet non-differentiable sparsity function which constitutes a tight lower bound on the cardinality function. The novelty of this approach is that we cast the problem of minimizing the new sparsity function as a problem with an indefinite quadratic objective function. We present a numerical comparison to other sparsity encouraging penalty functions for several applications. Additionally, we apply the techniques developed to minimize an objective function with a truncated hinge loss function. We present highly competitive results for all of the applications.

PDF - PASCAL Members only - Requires Adobe Acrobat Reader or other PDF viewer.
EPrint Type:Article
Project Keyword:Project Keyword UNSPECIFIED
Subjects:Learning/Statistics & Optimisation
ID Code:2665
Deposited By:Ron Meir
Deposited On:22 November 2006