Deep Sparse Coding Model And Its Application

Recent years have witnessed the great success of sparse coding in many areas, in-cluding data mining, machine learning, and computer vision. Sparse coding providesa class of unsupervised algorithms for learning a set of over-complete basis function-s, allowing to reconstruct the original signal by linearly combining a small subset ofthe bases. A shortcoming of most existing sparse coding algorithms is that they needto do some sort of iterative minimization to inference the sparse representations fortest 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 inputto sparse representation during the training procedure, predictive sparse decomposi-tion (PSD) can naturally be used for out-of-sample extension. Hence, PSD has recent-ly 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, PSDmay not achieve satisfactory performance in real applications. This paper proposes anovel model, called denoising PSD (DPSD), for robust sparse coding. Experimentson real visual object recognition tasks show that DPSD can dramatically outperformPSD in real applications. In addition, deep DPSD and deep multi-path DPSD are alsoproposed which apply DPSD in deep models and can achieve higher performance.