Δίκτυα βαθιάς πίστης

Ράμμος, Δημήτριος Χ.

Deep belief networks

Master Thesis

Author

Ράμμος, Δημήτριος Χ.

Date

2014-09-30

Abstract

Firstly, this thesis is provided a general introduction on the pattern recognition and its evolution through time and are mentioned most of the areas (character recognition, computer - aided diagnosis, data mining, etc) in which the results of pattern recognition have brought great progress or are able to bring in the future. It is also analyzed the categories of learning, such as supervised, unsupervised and semi-supervised learning, as well as the educational methods used by each of these methods. In the second chapter, in which it is entered the main part of the thesis it is tried to make the reader understand the meaning of deep belief networks and their use, giving some specific definitions and analyzing their properties. In the last chapter, it is analyzed another separate category of deep belief networks, the so-called modular deep belief networks which are able not to forget features, elements and data that have been taught in the past without having losses when new data is entered in place the previous ones. It is analyzed the operation of these networks compared to simple deep belief networks and through some experiments that have been conducted and some appropriate figures, so that the reader has the opportunity to understand their meaning and function completely.