Multiresolution Deep Belief Networks

Motivated by the observation that coarse and fine resolutions of an image reveal different structures in the underlying visual phenomenon, we present a model based on the Deep Belief Network (DBN) which learns features from the multiscale representation of images. A Laplacian Pyramid is first constructed for each image. DBNs are then trained separately at each level of the pyramid. Finally, a top level RBM combines these DBNs into a single network we call the Multiresolution Deep Belief Network (MrDBN). Experiments show that MrDBNs generalize better than standard DBNs on NORB classification and TIMIT phone recognition. In the domain of generative learning, we demonstrate the superiority of MrDBNs at modeling face images.