You want to train a deep convolutional neural network for object recognition. Which of the following does NOT typically improve the prediction accuracy on the test set? A. Pretrain the model on a large dataset. B. Use dropout as a regularization. C. Increase the number of GPUs. D. Use data augmentation