> We get stuck in local maxima at each layer and the critical energy required to escape out of it would be too large.

ML solution - many restarts, add momentum and in the end maybe all local maxima are equally good after all.