Entropic optimum synthesis of multi-layered feed-forward ANNs

Optimization of the neural architecture is often critical to design an efficient and feasible solution, in particular when a VLSI implementation is considered. This paper proposes an original approach to the synthesis of multilayered feed-forward ANNs based on the analysis of the information quantity flowing through the network. A layer is described as an information filter which selects the relevant characteristics until the complete classification is performed. The basic incremental method, including the training supervised procedure, is derived to design optimum (or nearly-optimum) neural paradigms. A significant variant is also proposed to improve performances