In this paper, we introduce functional networks as a generalization and ext
ension of the standard neural networks in the sense that every problem that
can be solved by a neural network can also be formulated by a functional n
etwork. But, more importantly, we give examples of problems that cannot be
solved using neural networks but can be naturally formulated using function
al networks. Functional networks are defined as a collection of connected f
unctional units on a set of nodes. A functional unit or neuron connects inp
ut nodes to output nodes. The values of the output nodes are calculated fro
m the values of the input nodes by given functions of one or several argume
nts. The main differences with neural networks are that (a) the neural func
tions can be multivariate and can be different from neuron to neuron tin wh
ich case, no weights are necessary, because they subsume by the different f
unctions) and (b) the neuron outputs can be coupled, that is, coincident. T
his mathematical model of functional networks parallels printed circuit boa
rds with electronic components, thus giving an intuitive interpretation to
functional networks and an interesting and natural additional application.
The existence of functional units with common outputs leads to functional e
quations whose solution can lead to substantial simplification of the initi
al topology of the network and the neural functions involved. Two types of
functional networks (the one-layer and serial functional networks) are disc
ussed in detail. For the one-layer functional networks, a very simple simpl
ification algorithm is given. For the serial functional networks, systems o
f functional equations are obtained. The methods are illustrated by several
examples of applications. (C) 2000 John Wiley & Sons, Inc.