Denoising predictive sparse decomposition

Recent years have witnessed the great success of sparse coding in many areas, including data mining, machine learning, and computer vision. Sparse coding provides a class of unsupervised algorithms for learning a set of over-complete basis functions, allowing to reconstruct the original signal by linearly combining a small subset of the bases. A shortcoming of most existing sparse coding algorithms is that they need to do some sort of iterative minimization to inference the sparse representations for test points, which means that it´s not convenient for these algorithms to perform out-of-sample extension. By additionally training a non-linear regressor that maps input to sparse representation during the training procedure, predictive sparse decomposition (PSD) can naturally be used for out-of-sample extension. Hence, PSD has recently become one of the most famous learning algorithms for sparse coding. However, when the training set is not large enough to capture the variations of the sample, PSD may not achieve satisfactory performance in real applications. In this paper, we propose a novel model, called denoising PSD (DPSD), for robust sparse coding. Experiments on real visual object recognition tasks show that DPSD can dramatically outperform PSD in real applications.