## Resumen

The ℓ_{0} regularized optimization (ℓ_{0}-RO) problem is a nonconvex problem that is central to several applications such as sparse coding, dictionary learning, compressed sensing, etc. Iterative algorithms for ℓ_{0} - RO problem are only known to have local or subsequence convergence properties i.e. the solution is trapped in a saddle point or in an inferior local solution. Inspired by techniques used to improve the alternating optimization (AO) of nonconvex functions, we propose a simple yet effective two step iterative method to improve the solution to the ℓ_{0}RO problem. Given an initial solution, we first find the vanilla solution to ℓ_{0}RO via a descent method (in particular, Nesterov's accelerated gradient descent), to then estimate a new initial solution by using a scaled version of the dictionary involved in the ℓ_{0}-RO problem, considering only a reduced number of its atoms. Our proposed algorithm is empirically demonstrated to have the best tradeoff between accuracy and computation time, when compared to state-of-the-art algorithms. Furthermore, due to its structure, our proposed algorithm can be directly apply to the convolutional formulation of ℓ_{0}-RO.

Idioma original | Inglés |
---|---|

Título de la publicación alojada | 2018 IEEE 13th Image, Video, and Multidimensional Signal Processing Workshop, IVMSP 2018 - Proceedings |

Editorial | Institute of Electrical and Electronics Engineers Inc. |

ISBN (versión impresa) | 9781538609514 |

DOI | |

Estado | Publicada - 27 ago. 2018 |

Evento | 13th IEEE Image, Video, and Multidimensional Signal Processing Workshop, IVMSP 2018 - Zagori, Grecia Duración: 10 jun. 2018 → 12 jun. 2018 |

### Serie de la publicación

Nombre | 2018 IEEE 13th Image, Video, and Multidimensional Signal Processing Workshop, IVMSP 2018 - Proceedings |
---|

### Conferencia

Conferencia | 13th IEEE Image, Video, and Multidimensional Signal Processing Workshop, IVMSP 2018 |
---|---|

País/Territorio | Grecia |

Ciudad | Zagori |

Período | 10/06/18 → 12/06/18 |

## Huella

Profundice en los temas de investigación de 'Improved Solution to the ℓ_{0}Regularized Optimization Problem via Dictionary-Reduced Initial Guess'. En conjunto forman una huella única.