Efficient highly over-complete sparse coding using a mixture model

摘要: Sparse coding of sensory data has recently attracted notable attention in research learning useful features from the unlabeled data. Empirical studies show that mapping into a significantly higher-dimensional space with sparse can lead to superior classification performance. However, computationally it is challenging learn set highly over-complete dictionary bases and encode test learned bases. In this paper, we describe mixture model produce high-dimensional representations very efficiently. Besides computational advantage, effectively encourages are similar each other enjoy representations. What's more, proposed be regarded as an approximation local coordinate (LCC), which states approximately nonlinear manifold locally linear manner. Therefore, feature by works pretty well classifiers. We apply PASCAL VOC 2007 2009 datasets for task, both achieving state-of-the-art performances.