The Curse Of Dimensionality And Gradient-Based Training Of Neural Networks: Shrinking The Gap Between Theory And Applications