Abstract
We introduce a sparse scattering deep convolutional neural network, whichprovides a simple model to analyze properties of deep representation learningfor classification. Learning a single dictionary matrix with a classifieryields a higher classification accuracy than AlexNet over the ImageNetILSVRC2012 dataset. The network first applies a scattering transform whichlinearizes variabilities due to geometric transformations such as translationsand small deformations. A sparse l1 dictionary coding reduces intra-classvariability while preserving class separation through projections over unionsof linear spaces. It is implemented in a deep convolutional network with ahomotopy algorithm having an exponential convergence. A convergence proof isgiven in a general framework including ALISTA. Classification results areanalyzed over ImageNet.