Sunday, 25th May 2014 - Preliminary timetable
|Time||Room 1||Room 2||Room 3|
|10h30-13h00||Sparsity 1||Stats & R|
|14h00-17h00||Sparsity 2||Stats & R||Bayesian|
Sparsity 1: Blind Source SeparationTutor: Jérôme Bobin
BackgroundThis tutorial will give essential insights into the use of blind source separation methods in astrophysics with a particular application in cosmology. Multichannel or multiwavelengths data are more than common in the field of astronomy and astrophysics. Analyzing this particular multidimensional data requires specific tools; these data are commonly modeled as the superposition of elementary components or sources which generally do not contribute similarly in each observation channel. The aim of blind source separation is to estimate simultaneously the sources and their contribution weight in each of these channels. Think for instance of a symphonic concert recorded with different microphones (i.e. the observation channels) and imagine the problem of separating out the contribution of each individual instrument from this cacophony. This tutorial will review the basics of blind source separation as well as more recently introduced sparsity-based techniques. An application to WMAP data will also be at the menu.
We invite all participants to bring their own laptop with Python and the following modules installed : scipy, numpy, matplotlib and pyfits.
Sparsity 2: Sparse regularization of inverse problemsTutor: François Lanusse
BackgroundIn this tutorial we will introduce powerful sparse regularization techniques to address ill-posed linear inverse problems. For such problems, the inverse operator is potentially unbounded and there is no unique solution in the absence of prior. As an example of this class of problems we will be introducing the specific case of deconvolution which is a recurring problem in signal processing and has been extensively covered for the past 30 years.
The aim of this tutorial is to have a hands on experience with sparse regularization techniques and with the selection of appropriate dictionaries to address the deconvolution problem. Several practical problems will be proposed illustrating the strength of the sparsity prior.
We expect all participants to bring their own laptop and we will provide a virtual machine in advance to ensure that all participants have a working environment with the necessary tools and test data sets before the workshop starts.
Statistical methodology and RTutor: Eric Feigelson
BackgroundThe fields of astronomy and statistics were, in past centuries, essentially merged but they diverged in the 20th century. The result today is that astronomers are poorly informed about the wealth of powerful methodologies developed by statisticians, and statisticians know little about contemporary challenges in modern astronomy. Statistics is needed for: understanding astronomical images, spectra and lightcurves; inference about underlying populations from limited samples; linking astronomical observations to astrophysical theories; and more. Fortunately, a very large, integrated and user friendly public domain software system has emerged in recent years to implement modern methods. R with its >5000 add-on CRAN packages has ~100,000 statistical functionalities with extensive graphics, links to other languages, and more. Over 100 textbooks and extensive on-line support provide guidance for the sophisticated R user.
The tutorial will be in five segments:
We expect all participants to bring their own laptop with R installed. R binary downloads are available for MacOS, Linux and Windows at http://www.r-project.org. CRAN packages and astronomical datasets are downloaded on-the-fly during the tutorials. R scripts will be available for cut-and-paste into the real-time R console from http://www2.astro.psu.edu/users/edf/SCCC21_May2014/ .
Bayesian CosmologyTutors: Daniel Mortlock and Andrew Jaffe
BackgroundMost researchers will at some point be required to perform some form of data analysis. This may be anything from simple line-fitting, through parameter estimation, to complex and computationally-demanding sampling for model selection on large datasets. Bayesian inference is a practical methodology for solving such problems that is built up from the fundamentals of probability theory.
The purpose of this short course is to provide understanding of Bayesian data analysis, especially in a cosmological context, concentrating on the underpinnings of probability theory, Bayes' theorem itself, and its application to the estimation of parameters. To apply these ideas, the tutorial will include discussion problems and a hands-on workshop implementing Monte Carlo Markov Chain (MCMC) samplers.
PreparationWe expect all participants to bring their own laptop, and we will provide a simple computational exercise in advance (in whatever language suits) to ensure they have appropriate software in place (including graph plotting) before the workshop starts.