A neural network for signal decomposition problems

This paper present the design of a neural network for signal decomposition problems with application examples. For this class of problems the proposed network has the same dynamics as the Hopfield net, but it is shown to realize the O(M2) connection paths among the M neurons with a number of wires and conductances increasing only linearly with increasing M, i.e. reducing this number by one dimension with respect to the quadratically increasing number of wires and conductances required in the Hopfield net. Other advantages of the proposed neural network are discussed in relation to classical examples of decomposition problems. In particular, a new architecture for an N-bit A/D converter is presented employing 4N conductances instead of the N(N + 1) Hopfield A/D conductances.