An algorithm for predicting the steady state behavior of binary CNNs

Stable CNN with binary outputs are used for real-time image processing. A fundamental step for CNN design is to develop simple and effective algorithms for predicting their steady-state behavior. So far, such algorithms are mainly based on the application of local rules, i.e. on the evaluation of the first order derivative of each cell at t = 0. They allow one to rigorously design only a small subset of templates. We have recently shown that the steady-state prediction can be improved by considering higher order derivatives at t = 0. In this manuscript we extend our previous results and show that the dynamic evolution of binary CNN can be accurately predicted through an algorithm, based on the evaluation of first order derivative zeros, when the latter are approximated through their Taylor expansion.