MNIST handwritten digit recognition with LeNet-1 CNN

This is an implementation of the method described in LeCun's 1989 "Handwritten Digit Recognition with a Back-Propagation Network" paper.

This is an implementation of the method described in LeCun's 1989 classic "Handwritten Digit Recognition with a Back-Propagation Network". It doesn't use any toolbox or library. Please note this was made as an exercise for me to grasp Convolutional Neural Networks, and is not meant to be computationnally efficient whatsoever. The training script takes 4 hours to complete one epoch on my laptop (i5 with 4GB of RAM). However, the first epochs already gives 94% good recognition, and I've gotten 97.03% after 4 epochs. It could turn out to be even higher, as I've discovered a very stupid mistake upon reviewing this code to post it here.