Compact representations and pruning in residual networks

Fereshteh Lagzi, Tonio Ball, Joschka Boedecker

Abstract: We show that residual networks encode their input signals in the transient dynamics of the neurons in each layer. These representations are similar for inputs from the same class, and distinct for inputs from different classes. Based on the neural transient dynamics, we provide a sufficient criterion to determine the depth of such networks during training. This criterion is based on the convergence of the neural dynamics in the last two successive layers of the residual block. This method compresses the depth of the network and removes unnecessary deep layers.

Keywords: Residual Networks, Dynamical Systems, Input Representations, Pruning