Learning Low-Rank Models From Compressive Measurements for Efficient Projection Design

Recent research has uncovered information-theoretic means to design projection matrices in scenarios where one information source is compressively sampled in the presence of a secondary source. Furthermore, if both sources can be approximated by Gaussian mixture (GM) models, it has been shown that it is possible to learn the characteristics of the secondary source from compressive measurements only. In this work, we investigate techniques that exploit low-rank GM approximations to the true distributions to reduce computational complexity and memory requirements during source learning and projection design. Two novel alternative projection design strategies are also introduced. These are tested against an existing strategy to determine which approach is superior for low size, weight, and power (SWAP) applications. Experimental results validate the benefits of the proposed low-rank strategies and reveal that all projection design algorithms offer similar levels of performance.