Approximation Property of Dropout Neural Networks: Sobolev Rates and Confidence Bounds

The universal approximation property of dropout neural networks does not by itself describe the network size required for an accurate random realization.