École Nationale Supérieure de
Techniques Avancées

Abstract

This internship was aiming at setting up an artificial intelligence based on neural networks able to recognize handwritten numerals and in analyzing their behaviour under various parameters. The MNIST dataset has been used to provide training and verification materials to networks. Two types of structures were studied; the first one is composed of fully-connected layers of neurons. The second one is called convolutional network with its main difference being analysing images through smaller parts. Then, different parameters were tested and performances were compared and analyzed in order to have a better understanding of neural networks behaviour. These experiments and analysis resulted in a knowledge source which could be used to apply neural networks to other problems.