Biologically inspired architecture of feedforward networks for signal classification

Abstract

The hypothesis is that in the lowest bidden layers of biological systems "local subnetworks" are smoothing an input signal. The smoothing accuracy may serve as a feature to feed the subsequent layers of the pattern classification network. The present paper suggests a multistage supervised and "unsupervised" training approach for design and training of multilayer feed-forward networks. Following to the methodology used in the statistical pattern recognition systems we split functionally the decision making process into two stages. In an initial stage, we smooth the input signal in a number of different ways and, in the second stage, we use the smoothing accuracy as anew feature to perform a final classification.