Imagine training a deep network where, as you move backward through the layers, the learning signals just get tinier and tinier, almost disappearing into thin air. This makes it super…