Network Architecture of Unsupervised Boltzmann Machines

Our thesis wants to illustrate recent developments in ANN, and study the topological properties of a specific type of ANN using tools from graph theory. The work is divided in two main parts. First, it presents useful concepts and models. We then focus on understanding the mode of operation of a Deep Belief Network (DBN), a multi-layer neural network that works under the unsupervised learning framework. The second part of this work analyzes a trained DBN (qualified on reading digits images from the popular MNIST database ADD REF) from a network perspective. We inspect the topological properties of the DBN, making use of graph theory. The goal of this unprecedented analysis is to seek a deeper knowledge of the topological modifications that the DBN experiences during the training.