Harmonic Analysis and Signal Processing Seminar
Iterative
Shrinkage/Thresholding Algorithms:
Some History and Recent Developments
Mario A. T. Figueiredo
Instituto de Telecomunicações
Instituto Superior Tecnico,
Lisboa, Portugal
Friday, March 6, 2009, 1:00pm, WWH 1314
Abstract
Iterative shrinkage/thresholding (IST) algorithms
are an important component of the computational toolbox used to address
linear inverse problems under sparseness-inducing regularization.
Examples of this class of problems include (among many others)
wavelet-based and total-variation-based image restoration, compressive
sensing, and coded aperture imaging. IST algorithms are typically used
to address unconstrained minimization problems where the objective
function includes a quadratic data term (for linear-Gaussian
observations) and a sparseness-inducing regularizer (typically a
1-norm). Via the recently proposed Bregman iterative approach,
IST algorithms also play a central role in dealing with the related
constrained optimization formulations. The purpose of this talk is
threefold: (a) To review the several (at least 4) different
perspectives from which IST algorithms can be built/derived, as well as
several convergence results. (b) To present recent advances, namely
several fast variants of IST which are the current state-of-the-art in
solving L2-L1 (and related) problems. (c) To describe recent
extensions, such as those designed for non-Gaussian noise and for
multiple regularizers.