Inverse Neural Networks

An Inverse Hollis-Paulos Artificial Neural Network


Louiza Sellami and Robert W. Newcomb


Abstract

The Hollis-Paulos artificial neural network (HPANN) is convenient in terms of its possibility for realization of variable weight ANNs in VLSI by MOS transistor circuits, though it is non-dynamical and not driven by external inputs. Here we introduce dynamics and inputs into the HPANN and show that over the range of operation covered by the Hollis-Paulos theory the system has an inverse. In particular we derive that inverse and give simulation results on its operation, showing how well the input to the original HPANN can be recovered from the output of the HPANN when fed into the inverse system. A comparison is made with the previous inverse of the Hopfield ANN. Possible applications of these inverse systems are to decoding of transmitted ANN signals and to inverse filtering for the extraction of input signals from processed signals.