A fully recursive perceptron network architecture

This paper introduces a fully recursive perceptron network (FRPN) architecture as a possible replacement of multilayer perceptron (MLP) networks. The FRPN consists of an input layer, an output layer, and only one hidden layer in which the hidden layer neurons are fully connected with algebraic (instantaneous) connections, and not delayed connections. It is shown that the FRPN has a computational capability exceeding that of MLPs. The FRPN is particularly attractive as an alternative to deep learning methods that use MLPs with multiple hidden layers since the FRPN eliminates the need of obtaining the number of layers and the number of neurons per layer. Some insight into the mechanisms of working of the FRPN is obtained through an application to a practical learning problem, viz., the handwritten digit recognition problem.