Performing signal processing tasks on compressive measurements of data hasreceived great attention in recent years. In this paper, we extend previouswork on compressive dictionary learning by showing that more general randomprojections may be used, including sparse ones. More precisely, we examinecompressive K-means clustering as a special case of compressive dictionarylearning and give theoretical guarantees for its performance for a very generalclass of random projections. We then propose a memory and computation efficientdictionary learning algorithm, specifically designed for analyzing largevolumes of high-dimensional data, which learns the dictionary from very sparserandom projections. Experimental results demonstrate that our approach allowsfor reduction of computational complexity and memory/data access, withcontrollable loss in accuracy.
展开▼
