Previously, for our digit problem, we used symmetry and intensity. How do these features relate to deep networks? Do we still need them?
Symmetry and intensity features can be the input features in the input layer or some outputs of some hidden layer. We may intentionally inject them into the network or they may present in the network unintentionally through training process. I think which choice to make depends on the problem we face. Personally, I do not prefer to take symmetry and intensity as the only features in the input layer as there are so many informations lost.