This doesn't hold for the networks currently in use, only if we want to more closely simulate human brains. There is no real indication yet that we can train these better or that they work better.