Contractive Slab and Spike Convolutional Deep Belief Network

Convolutional Deep Belief Network (CDBN) is typically classified into deep generative model. Although CDBN has demonstrated the powerful capacity of feature extraction in unsupervised learning, there still remain diverse challenges in the robust and high-quality feature extraction. This paper designs an advanced hierarchical generative model in order to tackle with these troubles. First, we modify conventional Convolutional Restricted Boltzmann Machine (CRBM) through inducing Gaussian hidden units subsequently following point-wise multiplication with the original binary spike hidden units for high-order feature extraction of the local patch. We theoretically derive entire inferences of this novel model. Second, we attempt to learn more robust features by minimizing L₂ norm of the jacobian of the extracted features producing from the modified model as novel regularization trick. This can introduce a localized space contraction benefit for robust feature extraction in turn. Finally, this paper construct a novel deep generative model, Contractive Slab and Spike Convolutional Deep Belief Network (CssCDBN), based on the modified CRBM, in order to learn deeper and more abstract features. The performances on diverse visual tasks indicate that CssCDBN is a more powerful model achieving impressive results over many currently excellent models.