Qualitative analysis of the BP composed of product units and summing units

In this paper, we qualitatively analyze networks that contain product units. By replacing the neurons in traditional backpropagation (BP) nets with product units in hidden layer gives us a different type of BP network called P-S model. We further extend P-S to P-S(in) by adding direct connections from input neurons to output neurons. By comparing with traditional BP nets that consists of ordinary summing units, we examine performance of product unit networks in solving TC, XOR, AOX, and other hard binary problems such as odd and even parity problems. The results show that product units outperforms traditional BP nets in terms of both hardware efficiency and training requirement.