This network takes a 28x28 MNIST image and crops a random 24x24 window before training on it (this technique is called data augmentation and improves generalization). Similarly to do prediction, 4 random crops are sampled and the probabilities across all crops are averaged to produce final predictions. The network runs at about 5ms for both forward and backward pass on my reasonably decent Ubuntu+Chrome machine.
By default, in this demo we're using Adadelta which is one of per-parameter adaptive step size methods, so we don't have to worry about changing learning rates or momentum over time. However, I still included the text fields for changing these if you'd like to play around with SGD+Momentum trainer.
Report questions/bugs/suggestions to @karpathy.