Sparse coding is used to describe the image features perceived in the human visual system. Sparse representation is a reasonable and effective representation of image features. We therefore introduce the restricted Boltzmann machine (RBM) to deep learning because of its reliable unsupervised ability to learn image features. Stacked sparse restricted Boltzmann machines (SRBMs) cannot only mimic the hierarchical organization of the cortex but also achieve more abstractive image features. So using the SRBM to obtain the sparse representation of image features attracts more attention in the field of AI. We introduce the basics of the RBM, describe its advantages and review thoroughly the existing work. Finally, we summarize open questions suggested in the last section and the future development.