In this paper we ask: given a training set typical of the signals we wish to measure, what are the optimal set of linear projections for compressed sensing ? We show that the optimal projections are in general not the principal components nor the independent components of the data, but rather a seemingly novel set of projections that capture what is still uncertain about the signal, given the training set. We also present experimental results indicating that the projections onto the learned {\em uncertain components} may far outperform random projections. This is particularly true in the case of natural images, where random projections have vanishingly small signal to noise ratio.
Joint work with Hyun-Sung Chan and Bill Freeman