While not strictly necessary, those breakthrough definitely helped: dropout and greedy layer-wise pretraining. Also convolutions are not only used in computer

While not strictly necessary, those breakthrough definitely helped: dropout and greedy layer-wise pretraining. Also convolutions are not only used in computer vision. For example, alphago used them: (the paper is called "Mastering the Game of Go with Deep Neural Networks and Tree Search"). In my opinion, I would say that convolutions should be useful whenever your data has a spatial aspect to it.