Correct implementation of digital neural networks

This paper presents a theoretical analysis and some experimental results concerning the effects of both weights and activation function actual implementation, for feedforward multi-layer digital neural networks. The analysis allows us to derive some practical boundaries on the number of bits for the weight representation and on the number of levels for a stepwise approximation of the activation function (however, the analysis can be applied to different implementation technologies). By following such a rules, the equivalence between the abstract neural model and an actual cost-effective VLSI implementation is a priori guaranteed