We show that the error of iteratively magnitude-pruned networks empirically
follows a scaling law with interpretable coefficients that depend on the
architecture and task. We functionally approximate the error of the pruned
networks, showing it is predictable in terms of an invariant t