A pruning method for multiple heterogeneous output neural networks

Abstract

A new complete procedure for the selection of pruning threshold in MIMO (multiple input multiple output) feedforward artificial neural networks (FANN) is presented. It is based on the evaluation of a local sensitivity index calculated with respect of any single output of the network. Special emphasis is given to a particular class of neural networks with multiple heterogeneous outputs. It will be shown how to take into account of the non-homogeneous nature of the outputs by deriving an ldquoimportance indexrdquo from the nonlinear correlation of data. An example of the proposed method will be shown by the development of a neural architecture devoted to a specific multi-output inversion system.