How many degrees of freedom do we need to train deep networks: a loss landscape perspective