Power series analyses of back-propagation neural networks

Presents a technique for analyzing backpropagation neural networks. Each hidden unit in the network is modeled as a power series of the net function. This approach allows determination of the degree of the overall polynomial discriminant, which approximates the network, potentially revealing the complexity of the decision boundary for the training data. Hidden units whose models are constant or have degree 1 can be pruned, thereby simplifying the network. The modeling technique can be used as a probe to investigate the success or failure of training. The approximation was applied to two example neural nets designed to perform nonlinear filtering tasks