We introduce a new regularization technique column drop which uses inner structure of CNNs for classification to make its output invariant to random crops of input image. Use of this regularization eliminates need in data ...