Finally using mnist.
Using the typical 784x15x10 architecture. Batch size = 100.
This sample should get a success rate of a 91% in the training set and a 94% in the test set
The numbers are an image 28x28, the data is arranged in array of 28x28=784 and encoded in a png file.
Credits for the png files (and the inspiration and lectures :) ) go to Andrej Karpathy