Upper bounds for robustness of CNN templates and a design approach for robust templates

CNNs constitute a class of spatially discrete, nonlinear dynamic systems. Once the inputs and the states are initialized, the dynamic of a CNN is determined by a set of parameters, so-called templates. We investigate issues concerning the dynamic behavior of a CNN due to variations in template values. In particular, we derive, based on the output invariance at the equilibrium, upper bounds for these variations. Furthermore, a general design approach for robust templates is proposed.