Comparison of defect compensation methods for feedforward neural networks

Recently, many defect compensation methods have been proposed for feedforward neural networks implemented in hardware devices. However, there are few accurate quantitative comparisons with the performance of these defect compensation methods. In this paper, we compare the following three defect compensation methods; partial retraining (PR) scheme, whole network backpropagation (BP) retraining and FT (fault-tolerant) BP method. The BP algorithm and PR scheme retrain the neural network after defects have occurred. The FTBP method tries to obtain the weights those are robust for the defects. We can say that both the BP algorithm and PR scheme are cure-type compensation methods and the FTBP method is a precaution-type compensation method. We compare the average recognition rate, average training time and the generalization ability among these three methods in detail. The experiments show that the whole network retraining by the BP algorithm has the highest reliability on the XOR problem and face image recognition problem on the neural networks with a single broken link defect and two broken link defects.