In this paper, a method for pruning hidden neurones is presented and illustrated on voiced-unvoiced-silence detection problem. It is based on the statistical study of the derivatives of the outputs of the model with regards to each hidden neurone. We claim that if the model is not using a particular neurone to estimate its outputs, then the corresponding sensitivities will have a low degree of significance. The aim of this analysis is the selection of an appropriate subset of neurones for each layer, to finally obtain a more parsimonious model.
F. Albu, A. Mateescu, N. Dumitriu, "Architecture Selection for a Multilayer Feedforward Network", International Conference on Microelectronics and Computer Science, Chishinau, Republic of Moldavia, pp. 131-134, October 1997