DOLFIN-digit online for integration neural networks

Abstract

In this paper we describe an approach for using digit online arithmetic in the field of neural network computation. Digit online, a serial most significant digit first arithmetic, shows significant advantages over all other digital implementations. The serial communication between the online modules make the implementation of connection intensive networks feasible. The accuracy of the computation is only loosely coupled with the chosen digit level range, which determine the necessary count of interconnections. Furthermore, the accuracy is eligible through the length of the processed digit vector. The goal of this paper is to develop a strategy for the implementation of different network models. The comparison with the results of other implementations illustrate the advantages of the digit online approaches and the suitability for the application in the field of neural networks.