Seminar afternoon of the Biomedical and Astronomical Signal Processing group

What? 3 seminars on inverse problems and imaging applications

When and where? Tuesday March 17 from 14.00 to 17.30 at Heriot Watt Earl Mountbatten Building EM 2.33.

Who? Our international speakers are the selected candidates for 3 recently opened postdoctoral positions with BASP@Edinburgh.

You will find the seminar schedule below. Feel free to attend part of the event only.

Coffee will be served between the seminars, and sparkling wine will close the event.

Hoping to see you all.

============================================================================
14.00 - 14.50 Dr Alexandru Onose

Title: Adaptive algorithms for sparse representations

Abstract: A vector or matrix is said to be sparse if the number of nonzero elements is significantly smaller than the number of zero elements.
Solving an estimation problem with prior knowledge of this underlying sparsity property, generally substantially improves the accuracy of the solution or even allows solutions to problems unfeasible by standard approaches.
With applications in many areas of signal processing, sparsity aware algorithms have come recently into focus as a better alternative in place of more established methods.

Finding a sparse linear estimator in an adaptive setup where the data are available sequentially in time, is generally more involved than in the batch context where the data are processed as a whole.
The adaptive algorithms generally have to be very computationally efficient since many problems require real time processing.
Further more, they are required to provide robust performance and to be able to track any changes in the parameter values or sparsity properties.

To generate such fast algorithms, one possibility is the use of randomized coordinate descent.
Such approach generally involves a probabilistic selection of coordinates to update, followed by the minimization of a constrained least squares criterion.
Two algorithms are proposed.
The first selects the support of the nonzero values sequentially, in a greedy fashion followed by the minimization of the least squares criterion on the support.
The second minimizes an $\ell_1$ penalized least squares criterion.
Both methods offer robust performance with low computational cost and propose easy tunable configuration parameters.

Bio: Alexandru Onose received the Diploma Engineer degree from Politehnica University of Bucharest, Romania, in 2009, and the Doctor of Science in Technology degree from Tampere University of Technology, Finland, in 2014.
He has been working as a researcher in the Signal Processing department of Tampere University of Technology where he studied sparsity aware algorithms with applications to adaptive signal processing.
His research interests include adaptive signal processing, sparse representation, convex optimization, and numerical methods and algorithms for signal processing.
============================================================================

14.50-15.10 Coffee

============================================================================
15.10 - 16.00 Mrs Arwa Dabbech

Title: Image reconstruction of radio interferometric data using sparse representations.
Abstract: The observed image of the sky through a radio telescope is the true sky blurred with the instrument's response and contaminated by noise. The imaging problem is an ill-posed inverse problem due to the sub-sampling in the Fourier domain of the radio measurements and has an infinite number of solutions. I will talk about a new radio deconvolution algorithm named MORESANE based on highly redundant and shift invariant dictionaries (The Isotropic Undecimated Wavelet Transform dictionaries). The algorithm is greedy and iterative, combining complementary types of sparse recovery methods; a synthesis approach is used for reconstructing images, in which the synthesis atoms, representing the unknown astronomical sources, are learned using analysis structured-sparsity priors. I will present the promising results of the deconvolution method applied on realistic simulations of radio images along with comparisons with standard and recent algorithms.
Bio: Arwa Dabbech received her National Engineer Diploma majoring in signals and systems from Tunisia Polytechnic School in 2011. She is part of the Galaxies and Cosmology team, in the framework of her Ph.D project at the Laboratoire Lagrange in the Observatoire de La Cote d’Azur, Nice, France, since 2011. She currently works in astronomical imaging under the supervision of Chiara Ferrari and Eric Slezak, in close collaboration with David Mary from the Signal Processing team of the same laboratory. Her Ph.D thesis focuses on the development of novel deconvolution algorithms for radio interferometric images dedicated to the exploitation of the SKA and its precursors.
============================================================================

16.00 - 16.20 Coffee

============================================================================
16.20 - 17.10 Dr Silvia Gazzola

Title: Topics in Iterative Regularization

Abstract: This talk is concerned with regularization methods for linear inverse problems. Inverse problems are ubiquitous in many areas of Science and Engineering; we mainly consider image restoration problems. Regularization must be employed in order to recover a meaningful approximation of the exact solution, i.e., the original system has to be replaced by a nearby problem with better numerical properties. During this talk we introduce some standard iterative regularization methods, which are suitable for large-scale problems, and we focus on Krylov subspace methods. We also present the recent class of Krylov-Tikhonov methods, which merge the direct and iterative approaches to regularization. We address new strategies to define regularization parameters and regularization matrices in the Krylov-Tikhonov framework; in particular, we explain how sparse reconstructions can be recovered. We provide theoretical insight, and we display the results of many realistic numerical experiments.

Bio: Dr Gazzola received the Ph.D. degree in Computational Mathematics from the University of Padova (Italy) in 2014. She is currently a Post Doctoral fellow at the University of Padova. Her present research interests include regularization methods for inverse problems, imaging problems, and numerical linear algebra. She has active international collaborations, and she recently spent some semesters at Emory University (USA) and Kent Sate University (USA).
=============================================================================

17.10 Wine