Robust fault tolerant training of feedforward neural networks

A very robust fault tolerant algorithm based on the off-line backpropagation algorithm for training feedforward artificial neural networks is proposed. The effect of all possible single faulty hidden neurons is incorporated during each weight updating phase. This is in contrast with random selection of faulty hidden neurons introduced previously. Simulation results indicate that the new algorithm results in a very robust internal representation, An enhanced version of the algorithm which outperforms all existing algorithms in its ability to tolerate faults of different types is introduced, The enhanced version could result in such a robust internal representation that it can tolerate other fault types for which the network is not trained. A modified version of the algorithm which can tolerate the failure of any pair of hidden neurons is also introduced and analyzed