Arithmetic for a high-speed adaptive learning network element

Abstract

This paper presents a novel arithmetic scheme for a high-speed adaptive learning network (ALN) element. An ALN is a self-organizing scheme for implementing the Kolmogorov-Gabor (K-G) polynomial which maps an input vector X into an output scalar Y. In the first layer of an ALN there are n(n-l) / 2 elements. In the next layer the number of elements needed depends upon the number of outputs that are propagated from the first layer. In this paper only the design of a single element is considered. An array of memories (RAMs) and a parallel adder are used to perform multinomial arithmetic for the element. The memory array contains subfunction values which are calculated by an external host computer and downloaded to the memory array. All the memories operate on the input variables concurrently via a common address bus. The subfunction values from the memory array are then summed by a parallel adder to obtain the output of the element. A complete ALN implemented with the proposed ALN elements has advantages in operation speed and less hardware.