Residual network with recurrent structures

We introduce a recurrent structure (spatially) on residual networks, which can improve the performance of the network while saving parameters. We investigate the behaviour of recurrent structures in residual networks based on Riemannian manifolds, introducing curvature as a metric for neural networks. We also experimentally verify that the gain due to the recurrent structure is related to the curvature, and demonstrate the generality of the recurrent structure as a method to improve the performance of the network.