Date: 30th January, 2008 at 1.00 pm
Venue: Video Conference Room EM1.27
Name: Dr Thomas Blumensath; Research Fellow
The acquisition, storage, transmission, processing and interpretation of signals often requires the assumption of signal structure. For example, Nyquist-Shannon sampling assumes signals to have a known bandwidth. In this talk, another powerful signal model is considered, the sparse signal model. These are quite general models that are, nevertheless, sufficiently constrained to allow novel solutions to many signal processing problems. The focus in this talk is on algorithmic aspects. Recently, we introduced the 'greedy' Gradient Pursuit framework, which allows sparse models to be used with relatively large datasets. In this talk, two extensions to this approach are considered. Firstly, because large is never large enough, we propose the use of an even greedier selection step. The second extension proposes the use of a Gradient Pursuit type algorithm to solve non-linear sparse models, which significantly widens the applicability of sparse signal modelling.