First a DOE report on the Mathematics for Analysis of Petascale Data. This report is a summary of a workshop that took place on June 3–5, 2008. Of note the following mention of Compressed Sensing:
5.4 Data and Dimension Reduction
Finding: Analysis of petascale data in their raw form is often infeasible. Instead, improved methods for data and dimension reduction are needed to extract pertinent subsets, features of interest, or low-dimensional patterns. As we have seen, many of the application domains important to the DOE are bedeviled by data that are high volume, high dimensional, or both.
Scalable, flexible data reduction. Many of the tools in the current mathematical analysis toolkit are extremely useful, but are inapplicable to petascale data because they do not scale. So if the tools can’t come to the data, then perhaps the data, suitably slimmed, can come to the tools. Accordingly, one identified need is for scalable tools for data reduction, for extracting relevant subsets, compressed representations, or just the features of interest.
One example is compressed sensing, a novel application of constrained l1 minimization that permits the exact reconstruction of sparse signals from what might have seemed to be incomplete measurements. Though promising, to be applicable to petabyte data it will require fast scalable algorithms for l1 minimization. Another example is feature extraction; most existing methods are serial, by implementation and often by nature.
Second, Raytheon is looking for an ATR Algorithm Design Engineer with a Ph.D and whose desired skill includes knowledge of compressive sensing.
No comments:
Post a Comment