Equalization of the Training Set For Backpropagation Networks Applied to Classification Problems

Título: Equalization of the Training Set For Backpropagation Networks Applied to Classification Problems

Autores: Liporace, Frederico dos Santos; Machado, Ricardo José; Barbosa, Valmir C.

Resumo: One of the problems faced by multilayer perceptrons trained by the backpropagation algorithm when applied to classification problems is the low sensitivity of the resulting network to classes statistically less represented in the training set. This paper proposes that is such cases it is better to build a modular network, assigning to each independent module the task of recognizing one specific class and rejecting the others. The use of a modular architecture enables the construction of a modified training set for each module that tries to minimize the problem of the less represented classes. This modification or equalization assings a relevance degree to each training sample. This gives each sample a different degree of importance, and has the same effect of the replication of some samples in the training set. This technique was applied in an application related to satellite imagery classification, and the obtained results show that the modules trained with the equalized training set exhibit far better results than others trained with the “plain” training set.

Palavras-chave: Neural networks; backpropagation; classification

Páginas: 5

Código DOI: 10.21528/CBRN1994-016

Artigo em PDF: CBRN1994-paper16.pdf

Arquivo BibTex: CBRN1994-paper16.bib