Marvin Lerousseau

 This post is a guide to a Python implementation of the paper “Best Practices for Convolutional Neural Networks Applied to Visual Document Analysis” published by a Microsoft Research team. After a bit of theory about convolutional neural networks, we will code a 98.2%-accuracy convolutional network to classify images from the MNIST dataset. Then, we will create functions to preprocess the dataset to enhance our classifier accuracy, up to 99.6%. This guide is intended for readers relatively at ease with Python and the major concepts of neural networks (including backpropagation).