Nonequilibrium training of Restricted Boltzmann Machines

Restricted Boltzmann machines (RBMs) are a very important unsupervised learning method in the machine learning research landscape. RBMs are frequently employed is the construction of generative models, which are very important for the development of neural networks. Unlike typical generative models, RBMs allow the generation of good quality samples by doing quick sampling. In addition, the choice of this learning method is due to its proximity to statistical physics. The goal of this thesis is to train RBMs and test, by observing the generated samples, their out of equilibrium (OOE) behavior. The dataset used is MNIST (represented by number images), trained on relatively simple RBMs. First, we evaluate the reproducibility of the model for a fixed number of training samples, then study the stability and variability of the model and the equilibrium, pseudo-equilibrium and OOE regimes by training the RBM with variable samples.