Behavior-driven minimal implementation of digital ANNs

A theoretical analysis and some experimental results concerning the minimal implementation of multilayer feed-forward special-purpose neurocomputer is here presented. The goal of the paper is to provide a deterministic methodology to investigate how the typical customizations, operating with finite-precision arithmetic for synaptic weights representation and activation function approximation, affect the network behavior. The presented analysis allows the determination of generally applicable practical boundaries on the number of bits to be used by the various units composing digital realization of neurons. By following such constraints, it is a priori guaranteed the adherence between the abstract neural model and its actual cost-effective VLSI implementation.