Feedforward neural networks with improved insensitivity abilities
Cesare Alippi · 2003
The paper studies the insensitivity of regression-type feedforward neural networks, i.e., the ability possessed by the model of providing a graceful loss in performance when affected by perturbations. Such ability is somehow related to the application and, in general, cannot be simply improved by acting on the obtained model with off-line transformations. The attention is focused an perturbations affecting the network's weights. We identify the worst case perturbation and quantify its effect on the network output. Modifications of the training function are suggested to improve the overall insensitivity of the model.