Randomized iterative hard thresholding for sparse approximations

Robert Crandall, Bin Dong, Ali Bilgin

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Typical greedy algorithms for sparse reconstruction problems, such as orthogonal matching pursuit and iterative thresholding, seek strictly sparse solutions. Recent work in the literature suggests that given a priori knowledge of the distribution of the sparse signal coefficients, better results can be obtained by a weighted averaging of several sparse solutions. Such a combination of solutions, while not strictly sparse, approximates an MMSE estimator and can outperform strictly sparse solvers in terms of l-2 reconstruction error. We introduce a novel method for obtaining such an approximate MMSE estimator by replacing the deterministic thresholding operator of Iterative Hard Thresholding with a randomized version. We demonstrate the improvement in performance experimentally for both synthetic 1D signals and real images.

Original languageEnglish (US)
Title of host publicationProceedings - DCC 2014
Subtitle of host publication2014 Data Compression Conference
PublisherInstitute of Electrical and Electronics Engineers Inc.
Number of pages1
ISBN (Print)9781479938827
DOIs
StatePublished - Jan 1 2014
Event2014 Data Compression Conference, DCC 2014 - Snowbird, UT, United States
Duration: Mar 26 2014Mar 28 2014

Publication series

NameData Compression Conference Proceedings
ISSN (Print)1068-0314

Other

Other2014 Data Compression Conference, DCC 2014
CountryUnited States
CitySnowbird, UT
Period3/26/143/28/14

ASJC Scopus subject areas

  • Computer Networks and Communications

Cite this

Crandall, R., Dong, B., & Bilgin, A. (2014). Randomized iterative hard thresholding for sparse approximations. In Proceedings - DCC 2014: 2014 Data Compression Conference [6824455] (Data Compression Conference Proceedings). Institute of Electrical and Electronics Engineers Inc.. https://doi.org/10.1109/DCC.2014.25