Estudo sobre arquiteturas, regularização e otimização de sistemas de aprendizado profundo.

Abstract

Due to the recent development of the area of deep learning in the field of artificial intelligence, many theoretical and practical concepts are still obscure and with little documentation for beginning researchers in the area. In this context, this undergraduate thesis aims to explain some basic concepts and their influences on the performance and results of deep learning systems and neural networks. In addition to the theory, didactic experiments were implemented to exemplify practical constructions of systems with different architectures, for different problems, comparing the activation functions, optimization and regularization algorithms, initial parameter values, among others, optimizing the networks with better design choices for the specific problems.