Digital Neural Network Architecture and Implementation

Abstract

Digital implementations of neural networks represent a mature and well understood technology, which offers greater flexibility, scalability, and accuracy than the analog implementations. For example, using digital logic and memory it is quite easy to partition a large problem so that it can be solved by a smaller (in terms of hardware) implementation. Using the same logic and memory, a digital implementation can realize more than one network and combine the results in a hierarchical fashion to solve large problems. The main drawbacks of digital VLSI implementations are however their larger silicon area, relatively slower speed and the great cost of interconnecting processing units. These problems are addressed in this paper and solutions to alleviate them are proposed. First, the theoretical foundations for digital VLSI implementations are developed. Based on those, the design of digital systems is proposed that offer great speed and at the same time do not require a large number of interconnections.