Deep Learning as a Convex Paradigm of Computation: Minimizing Circuit Size with ResNets