AINS: architecture independent neuron selection
Fabrice Bossaert, D. Heavner Benjamin · 2003
AINS, a new method to select relevant variables in the input of connectionist systems is presented. This method, based on a measure giving the contribution of an input neuron on an output one, allows one to identify and select important variables in a feature space. The proposed approach is sufficiently general for applying to all the existing feedforward architectures like multilayer-perceptrons, as well as to those using Euclidean units like radial basis function networks. Experimental validation is shown with a difficult problem - noisy Breiman waveforms.