* The way modern CNNs get around this rotation problem is by augmenting the training data to include huge numbers of examples of the same object rotated.