Modul RBM a DBM pro program Modeler neuronových sítí

Abstract

This master thesis focuses on extending the program Neural net modeler with Restricted Boltzmann Machine and Deep Boltzmann Machine networks. Furthermore, this thesis deals with parallelization of learning of these networks. Stated neural networks and monitoring possibilities of their learning are described in this work. Possibilities of neural networks parallelization for distributed computing are also described there. In the scope of this study, the program Neural net modeler was extended and a new program was created, which deals with parallelization of Restricted Boltzmann Machine and Deep Boltzmann Machine networks using data parallelism. Using these programs and a supercomputer, the speed of learning and the success rate of pattern classification were compared between sequential and parallel implementation of the mentioned neural networks on various datasets.