In this paper an extension of HUMANN (hierarchical unsupervised modular adaptive neural network) is presented together with a parametric study of this network in dealing with noise and with classes of any shape and size. The study has been made based on the two most noise dependent HUMANN parameters, λ and μ, using synthesised databases (bidimensional patterns with outliers and classes with different probability density distribution). In order to evaluate the robustness of HUMANN a Monte Carlo [1] analysis was carried out using the creation of separate data in given classes. The influence of the different parameters in the recovery of these classes was then studied.