Título: As Redes Neurais de Hopfield e Multi-Layer Perceptrons Formando uma Arquitetura Neural Híbrida (MLP+H) com Características Próprias
Autores: Silva, Clayton Oliveira; Hernandez, Emílio Del Moral
Resumo: This article has the objective of presenting the first results of the project 1 of one hybrid neural network composed by the junction of two different types of intelligent computational systems: the Hopfield neural network and the Multi-Layer Perceptron. In this sense, it was possible to develop one hybrid neural system that presents better performance, in some cases, than the two cited types of neural networks. For example, the immunity to external noises added to the input patterns was improved. Additionally, this hybrid system has the ability of dealing with different types of data when compared with classical architectures. For example, while the Hopfield network deals with binary patterns, and the MLP produces one function that maps inputs to outputs by its training, dealing with information that have always analog characteristics, the hybrid network MLP+H works with analog inputs and digital outputs. As it will be discussed, the hybrid neural architecture developed in the project here presented has a good flexibility in various applications, as anti-noise filters, what is actually being studied. Also important in this paper is the discussion of concepts and methodologies of evaluation and characterization of performance of the considered architectures based on the confusion matrix concept, which allows a deeper analysis of experimental results.
Código DOI: 10.21528/CBRN2003-005
Artigo em PDF: 6CBRN_005.PDF
Arquivo BibTex: 6CBRN_005.bib