Learning algorithms for fault tolerance in radial basis function networks

This paper investigates the incorporation of fault tolerance at the learning stage into Radial Basis Function (RBF) networks. The approach is particularly attractive since the cost of fault detection and correction in a practical VLSI implementation of such networks could be prohibitive due to the large number of neurons and connections. The RBF networks considered are applied to the task of analog function approximation. A fairly general fault model is considered wherein faulty neurons are assumed to be stuck at a random value. Two new learning methods based on regression are proposed to learn the weights and one new regression based learning method is proposed to learn the centers. Simulation results are presented which show that a considerable improvement in fault tolerance can be achieved over the non-fault-tolerant learning algorithm