Performance evaluation of a novel fault tolerance training algorithm

This paper presents a performance evaluation of a novel algorithm for fault tolerance training of artificial neural networks (ANNs). The proposed algorithm is based on a genetic algorithms technique. A realistic, and practical fault model is adopted, it reflects the failures that arise during hardware realization of ANNs, regardless of the hardware platform used in the implementation. Using this fault model, an algorithm is developed and experimental results are performed to test the validity of the algorithm for different feedforward network sizes and types, and to check the ability of the algorithm to cover other fault models as a subset of the adopted one. A comparison with the conventional backpropagation learning algorithm is performed. The results show that the proposed algorithm is superior to the backpropagation from the fault tolerance point of view. The proposed algorithm has potential benefits in designing ANNs that can tolerate internal faults in the hardware realization of ANNs by incorporating fault tolerance in the training phase