Transputer based simulation of a general purpose, fault tolerant neural network

Abstract

Research on neural networks is to a large extent dependent upon the use of computer simulations. The availability of the Transputer hardware [1] and its accompanying OCCAM [2, 3] software together with the highly parallel nature of the neural algorithms has led us to develop a Transputer based simulation of a neural network.

The simulated model has deterministic activation function and syncronous updating (according to the classification of neural models proposed in [4]), and its processing elements are linear threshold units.

We look at the network as a parallel general purpose computer which can be programmed assigning the connection weights. To this end it has been developed an assembly programming language which we use to configure the network and to make it partially fault tolerant.

Supported in part through the MPI 40% fund and a contract with the CNR “Progetto finalizzato Informatica” and one with the “Progetto finalizzato Robotica”.