Sensitivity and learning of two digital artificial neural network structures

We extend the analysis of parameter sensitivity and interdependence to two digital artificial neural network structures, the backpropagation and ALCOVE. This paper compares the two networks, and we generalize to show that a highly sensitive weight contributes more to the prediction of the network than does an insensitive parameter. This suggests that the information structure of an input pattern can be determined by looking at the sensitivity of the interconnection weights, which has ramifications in network design. Additionally, results from a different set of simulations indicate that information about weight sensitivity and interdependence is predictive of the learning behavior of the networks