I've built a fully-connected feed-forward neural network to recognize handwritten digits. I used MNIST and another very similar dataset (containing Arabic digits - same training set and test set count as those of MNIST).
For a network with exactly the same architecture (1 hidden layer, same learning rate, same weight sampling), I tested on a sub-sample that used 100 images for training. The test accuracy was very low, but what I noticed was that for the Arabic digits, the network took significantly longer to train (double the amount of time than for English digits - 33ms and 15ms, respectively).
What could be the reason for this?