Sparse Representation Using Deep Learning to Classify Multi-Class Complex Data

Extracting best feature set to reinforce discrimination is always a challenge in machine learning. In this paper, a method named General Locally Linear Combination (GLLC) is proposed to extract automatic features using a deep autoencoder and also to reconstruct a sample based on the other samples sparsely in a low-dimensional space. Extracting features along with the discrimination ability of the sparse models have created a robust classifier that shows simultaneous reduction in samples and features. To enhance the capability of this scheme, some feature sets from several layers of an autoencoder are combined and an extension of GLLC has been proposed that called here as Multi-modal General Locally Linear Combination. Although the main application of the proposed methods is in visual classification and face recognition, they have been used in other applications. Extensive experiments are conducted to demonstrate that the proposed algorithms gain high accuracy on various datasets and outperform the rival methods.