Sparse coding in practice

The goal in sparse coding is to seek a linear basis representation where each image is represented by a small number of active coefficients. The learning algorithm involves adapting a basis vector set while imposing a low-entropy, or sparse, prior on the output coefficients. Sparse coding applied on natural images has been shown to extract wavelet-like structure [9, 4]. However, our experience in using sparse coding for extracting multi-scale structure in object-specific ensembles, such as face images or images of a gesturing hand, has been negative. In this paper we highlight three points about the reliability of sparse coding for extracting the desired structure: (1) using an overcomplete representation (2) projecting data into a low-dimensional subspace before attempting to resolve the sparse structure and (3) applying sparsity constraint on the basis elements, as opposed to the output coefficients.