A parametric study of humann in relation to the noise: application for the identification of compounds of environmental interest

Abstract

In this paper we present a parametric study of a hierarchical unsupervised modular adaptive neural network (HUMANN), in dealing with noise. HUMANN is a biologically plausible feedforward neural architecture which has the capacity for working in domains with noise and overlapping classes, with no priori information of the number of different classes in the data, with highly non-linear boundary class and with high dimensionality data vectors. It is appropriate for classification processes performing blind clustering. The study has been accomplished round the two most noise-dependent HUMANN parameters, λ and ρ, using synthesized databases (sinusoidal signals with Gaussian noise). We show that HUMANN is highly resistant to noise, improving the performance of different neural architectures such as ART2 and DIGNET. We also present the application of HUMANN for the identification of pollutants in the environment. Specifically it has been tested with Polychlorinated dibenzofurans (PCDFs), some of the most hazardous pollutants of the environment.